id
int64 1.9B
3.25B
| title
stringlengths 2
244
| state
stringclasses 2
values | body
stringlengths 3
58.6k
⌀ | created_at
timestamp[s]date 2023-09-15 14:23:33
2025-07-22 09:33:54
| updated_at
timestamp[s]date 2023-09-18 16:20:09
2025-07-22 10:44:03
| closed_at
timestamp[s]date 2023-09-18 16:20:09
2025-07-19 22:45:08
⌀ | html_url
stringlengths 49
51
| pull_request
dict | number
int64 6.24k
7.7k
| is_pull_request
bool 2
classes | comments
listlengths 0
24
|
|---|---|---|---|---|---|---|---|---|---|---|---|
3,120,799,626 |
Add option to ignore keys/columns when loading a dataset from jsonl(or any other data format)
|
open
|
### Feature request
Hi, I would like the option to ignore keys/columns when loading a dataset from files (e.g. jsonl).
### Motivation
I am working on a dataset which is built on jsonl. It seems the dataset is unclean and a column has different types in each row. I can't clean this or remove the column (It is not my data and it is too big for me to clean and save on my own hardware).
I would like the option to just ignore this column when using `load_dataset`, since i don't need it.
I tried to look if this is already possible but couldn't find a solution. if there is I would love some help. If it is not currently possible, I would love this feature
### Your contribution
I don't think I can help this time, unfortunately.
| 2025-06-05T11:12:45 | 2025-06-28T09:03:00 | null |
https://github.com/huggingface/datasets/issues/7594
| null | 7,594 | false |
[
"Good point, I'd be in favor of having the `columns` argument in `JsonConfig` (and the others) to align with `ParquetConfig` to let users choose which columns to load and ignore the rest",
"Is it possible to ignore columns when using parquet? ",
"Yes, you can pass `columns=...` to load_dataset to select which columns to load, and it is passed to `ParquetConfig` :)",
"Ok, i didn't know that. \nAnyway, it would be good to add this to others",
"Hi @lhoestq \n\nI'd like to take this up!\n\nAs you suggested, I’ll extend the support for the columns parameter (currently used in ParquetConfig) to JsonConfig as well. This will allow users to selectively load specific keys/columns from .jsonl (or .json) files and ignore the rest — solving the type inconsistency issues in unclean datasets.",
"Hi @avishaiElmakies and @lhoestq \n\nJust wanted to let you know that this is now implemented in #7594\nAs suggested, support for the `columns=...` argument (previously available for Parquet) has now been extended to **JSON and JSONL** loading via `load_dataset(...)`. You can now load only specific keys/columns and skip the rest — which should help in cases where some fields are unclean, inconsistent, or just unnecessary.\n\n### ✅ Example:\n\n```python\nfrom datasets import load_dataset\n\ndataset = load_dataset(\"json\", data_files=\"your_data.jsonl\", columns=[\"id\", \"title\"])\nprint(dataset[\"train\"].column_names)\n# Output: ['id', 'title']\n```\n\n### 🔧 Summary of changes:\n\n* Added `columns: Optional[List[str]]` to `JsonConfig`\n* Updated `_generate_tables()` to filter selected columns\n* Forwarded `columns` argument from `load_dataset()` to the config\n* Added test case to validate behavior\n\nLet me know if you'd like the same to be added for CSV or others as a follow-up — happy to help.",
"@ArjunJagdale this looks great! Thanks!\nI believe that every format that is supported by `datasets` should probably have this feature since it is very useful and will streamline the api (people will know that they can just use `columns` to select the columns they want, and it will not be dependent on the data format) ",
"Thanks @avishaiElmakies — totally agree, making `columns=...` support consistent across all formats would be really helpful for users."
] |
3,118,812,368 |
Fix broken link to albumentations
|
closed
|
A few months back I rewrote all docs at [https://albumentations.ai/docs](https://albumentations.ai/docs), and some pages changed their links.
In this PR fixed link to the most recent doc in Albumentations about bounding boxes and it's format.
Fix a few typos in the doc as well.
| 2025-06-04T19:00:13 | 2025-06-05T16:37:02 | 2025-06-05T16:36:32 |
https://github.com/huggingface/datasets/pull/7593
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7593",
"html_url": "https://github.com/huggingface/datasets/pull/7593",
"diff_url": "https://github.com/huggingface/datasets/pull/7593.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7593.patch",
"merged_at": "2025-06-05T16:36:32"
}
| 7,593 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7593). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.",
"@lhoestq ping"
] |
3,118,203,880 |
Remove scripts altogether
|
closed
|
TODO:
- [x] remplace fixtures based on script with no-script fixtures
- [x] windaube
| 2025-06-04T15:14:11 | 2025-07-16T18:59:07 | 2025-06-09T16:45:27 |
https://github.com/huggingface/datasets/pull/7592
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7592",
"html_url": "https://github.com/huggingface/datasets/pull/7592",
"diff_url": "https://github.com/huggingface/datasets/pull/7592.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7592.patch",
"merged_at": "2025-06-09T16:45:27"
}
| 7,592 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7592). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.",
"Hi @lhoestq,\r\nI wanted to ask\r\nare you planning to stop supporting dataset builds using `GeneratorBasedBuilder`?\r\n\r\nIf so, could you share the reason why?",
"We stopped supporting dataset scripts altogether, whether they are based on GeneratorBasedBuilder or any other builder. This means you can't `load_dataset()` a dataset script anymore. We did this mostly for security reasons which is blocking for many users and also impossible to build upon (e.g. the for the Dataset Viewer on HF)",
"Ah, so only the `trust_remote_code` feature of `load_dataset` is deprecated, and\r\n\r\n```python\r\nfrom datasets import load_dataset_builder\r\n \r\nbuilder = load_dataset_builder('cornell-movie-review-data/rotten_tomatoes') \r\nbuilder.download_and_prepare() \r\n```\r\n\r\nwe can still load data using `load_dataset_builder` and `download_and_prepare`, right?\r\nThat's a relief. I thought the removal of `trust_remote_code` in `load_dataset` meant `GeneratorBasedBuilder` was being deprecated too, haha.\r\nGot it, thanks for the clarification!\r\n"
] |
3,117,816,388 |
Add num_proc parameter to push_to_hub
|
open
|
### Feature request
A number of processes parameter to the dataset.push_to_hub method
### Motivation
Shards are currently uploaded serially which makes it slow for many shards, uploading can be done in parallel and much faster
| 2025-06-04T13:19:15 | 2025-06-27T06:13:54 | null |
https://github.com/huggingface/datasets/issues/7591
| null | 7,591 | false |
[
"Hi @SwayStar123 \n\nI'd be interested in taking this up. I plan to add a `num_proc` parameter to `push_to_hub()` and use parallel uploads for shards using `concurrent.futures`. Will explore whether `ThreadPoolExecutor` or `ProcessPoolExecutor` is more suitable based on current implementation. Let me know if that sounds good!\n",
"Just a quick update — `push_to_hub()` already had the `num_proc` argument in its signature and was correctly passing it internally to `_push_parquet_shards_to_hub()`.\n\nThe actual change required was inside `_push_parquet_shards_to_hub()` to enable parallel shard uploads using `multiprocessing` when `num_proc > 1`.\n\n@lhoestq @SwayStar123 ",
"> Hi @SwayStar123 \n> \n> I'd be interested in taking this up. I plan to add a `num_proc` parameter to `push_to_hub()` and use parallel uploads for shards using `concurrent.futures`. Will explore whether `ThreadPoolExecutor` or `ProcessPoolExecutor` is more suitable based on current implementation. Let me know if that sounds good!\n> \n\nHey thanks for working on it. But I'm not a hf dev so I don't know the best way to do it."
] |
3,101,654,892 |
`Sequence(Features(...))` causes PyArrow cast error in `load_dataset` despite correct schema.
|
closed
|
### Description
When loading a dataset with a field declared as a list of structs using `Sequence(Features(...))`, `load_dataset` incorrectly infers the field as a plain `struct<...>` instead of a `list<struct<...>>`. This leads to the following error:
```
ArrowNotImplementedError: Unsupported cast from list<item: struct<id: string, data: string>> to struct using function cast_struct
```
This occurs even when the `features` schema is explicitly provided and the dataset format supports nested structures natively (e.g., JSON, JSONL).
---
### Minimal Reproduction
[Colab Link.](https://colab.research.google.com/drive/1FZPQy6TP3jVd4B3mYKyfQaWNuOAvljUq?usp=sharing)
#### Dataset
```python
data = [
{
"list": [
{"id": "example1", "data": "text"},
]
},
]
```
#### Schema
```python
from datasets import Features, Sequence, Value
item = Features({
"id": Value("string"),
"data": Value("string"),
})
features = Features({
"list": Sequence(item),
})
```
---
### Tested File Formats
The same schema was tested across different formats:
| Format | Method | Result |
| --------- | --------------------------- | ------------------- |
| JSONL | `load_dataset("json", ...)` | Arrow cast error |
| JSON | `load_dataset("json", ...)` | Arrow cast error |
| In-memory | `Dataset.from_list(...)` | Works as expected |
The issue seems not to be in the schema or the data, but in how `load_dataset()` handles the `Sequence(Features(...))` pattern when parsing from files (specifically JSON and JSONL).
---
### Expected Behavior
If `features` is explicitly defined as:
```python
Features({"list": Sequence(Features({...}))})
```
Then the data should load correctly across all backends — including from JSON and JSONL — without any Arrow casting errors. This works correctly when loading from memory via `Dataset.from_list`.
---
### Environment
* `datasets`: 3.6.0
* `pyarrow`: 20.0.0
* Python: 3.12.10
* OS: Ubuntu 24.04.2 LTS
* Notebook: \[Colab test notebook available]
---
| 2025-05-29T22:53:36 | 2025-07-19T22:45:08 | 2025-07-19T22:45:08 |
https://github.com/huggingface/datasets/issues/7590
| null | 7,590 | false |
[
"Hi @lhoestq \n\nCould you help confirm whether this qualifies as a bug?\n\nIt looks like the issue stems from how `Sequence(Features(...))` is interpreted as a plain struct during schema inference, which leads to a mismatch when casting with PyArrow (especially with nested structs inside lists). From the description, this seems like an inconsistency with expected behavior.\n\nIf confirmed, I’d be happy to take a shot at investigating and potentially submitting a fix.\n\nAlso looping in @AHS-uni — could you kindly share a minimal JSONL example that reproduces this?\n\nThanks!",
"Hello @Flink-ddd \n\nI updated the minimal example and included both JSON and JSONL minimal examples in the Colab notebook. \n\nHere is the minimal JSON file for convenience (can't upload JSONL files).\n\n[mini.json](https://github.com/user-attachments/files/20535145/mini.json)\n\nI've also found a number of issues which describe a similar problem:\n\n[7569](https://github.com/huggingface/datasets/issues/7569) (Open)\n[7137](https://github.com/huggingface/datasets/issues/7137) (Open)\n[7501](https://github.com/huggingface/datasets/issues/7501) (Closed)\n[2434](https://github.com/huggingface/datasets/issues/2434) (Closed)\n\nThe closed issues don't really address the problem (IMO). [7501](https://github.com/huggingface/datasets/issues/7501) provides a workaround (using a Python list instead of `Sequence`), but it seem precarious. ",
"Hi ! `Sequence({...})` corresponds to a struct of lists ([docs](https://huggingface.co/docs/datasets/v3.6.0/en/package_reference/main_classes#datasets.Features)). This come from Tensorflow Datasets.\n\nIf you want to use a list of structs, you should use `[{...}]`, e.g.\n\n```python\nitem = {\n \"id\": Value(\"string\"),\n \"data\": Value(\"string\"),\n}\n\nfeatures = Features({\n \"list\": [item],\n})\n```",
"@lhoestq Thanks for your explanation, which helps me understand the logic behind. But I'm confused how to define that in `README.md`?\n\nMy jsonl data is: \n```\n{\"answers\": [{\"text\": \"text1\", \"label\": \"label1\"}, {\"text\": \"text2\", \"label\": \"label2\"},]}\n{\"answers\": [{\"text\": \"text1\", \"label\": \"label1\"}, {\"text\": \"text2\", \"label\": \"label2\"},]}\n...\n```\n\nMy README.md look like\n```\ndataset_info:\n- config_name: default\n features:\n - name: answers\n sequence:\n - name: text\n dtype: string\n - name: label\n dtype: string\n```\nI understand `sequence` here is not correct, but what's the correct format? I tried following (`sequence -> dtype`)and seems not the case:\n```\ndataset_info:\n- config_name: default\n features:\n - name: answers\n dtype:\n - name: text\n sequence: string\n - name: label\n sequence: string\n```",
"The `List` type which doesn't have the weird dict behavior of `Sequence` has been added for `datasets` 4.0 (to be released next week). Feel free to install `datasets` from source to try it out :)\nEDIT: it's out !\n\nYou can fix the issue using `List` instead of `Sequence`, e.g. in the case of the original post:\n\n```python\n# Feature spec with List of structs\nitem = {\n \"id\": Value(\"string\"),\n \"data\": Value(\"string\"),\n}\n\nfeatures = Features({\n \"list\": List(item),\n})\n```\n\nfor which the README.md is\n\n```yaml\ndataset_info:\n- config_name: default\n features:\n - name: list\n list:\n - name: id\n dtype: string\n - name: data\n dtype: string\n```",
"@lhoestq Thanks! I didn't realize there is a `list` keyword I could use. I thought I had to use `dtype` or something. Hope there could be better documentation on the `README.md` formats. I've closed my issue #7137 "
] |
3,101,119,704 |
feat: use content defined chunking
|
open
|
WIP:
- [x] set the parameters in `io.parquet.ParquetDatasetReader`
- [x] set the parameters in `arrow_writer.ParquetWriter`
It requires a new pyarrow pin ">=21.0.0" which is not yet released.
| 2025-05-29T18:19:41 | 2025-06-17T15:04:07 | null |
https://github.com/huggingface/datasets/pull/7589
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7589",
"html_url": "https://github.com/huggingface/datasets/pull/7589",
"diff_url": "https://github.com/huggingface/datasets/pull/7589.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7589.patch",
"merged_at": null
}
| 7,589 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7589). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.",
"Need to set `DEFAULT_MAX_BATCH_SIZE = 1024 * 1024`"
] |
3,094,012,025 |
ValueError: Invalid pattern: '**' can only be an entire path component [Colab]
|
closed
|
### Describe the bug
I have a dataset on HF [here](https://huggingface.co/datasets/kambale/luganda-english-parallel-corpus) that i've previously used to train a translation model [here](https://huggingface.co/kambale/pearl-11m-translate).
now i changed a few hyperparameters to increase number of tokens for the model, increase Transformer layers, and all
however, when i try to load the dataset, this error keeps coming up.. i have tried everything.. i have re-written the code a hundred times, and this keep coming up
### Steps to reproduce the bug
Imports:
```bash
!pip install datasets huggingface_hub fsspec
```
Python code:
```python
from datasets import load_dataset
HF_DATASET_NAME = "kambale/luganda-english-parallel-corpus"
# Load the dataset
try:
if not HF_DATASET_NAME or HF_DATASET_NAME == "YOUR_HF_DATASET_NAME":
raise ValueError(
"Please provide a valid Hugging Face dataset name."
)
dataset = load_dataset(HF_DATASET_NAME)
# Omitted code as the error happens on the line above
except ValueError as ve:
print(f"Configuration Error: {ve}")
raise
except Exception as e:
print(f"An error occurred while loading the dataset '{HF_DATASET_NAME}': {e}")
raise e
```
now, i have tried going through this [issue](https://github.com/huggingface/datasets/issues/6737) and nothing helps
### Expected behavior
loading the dataset successfully and perform splits (train, test, validation)
### Environment info
from the imports, i do not install specific versions of these libraries, so the latest or available version is installed
* `datasets` version: latest
* `Platform`: Google Colab
* `Hardware`: NVIDIA A100 GPU
* `Python` version: latest
* `huggingface_hub` version: latest
* `fsspec` version: latest
| 2025-05-27T13:46:05 | 2025-05-30T13:22:52 | 2025-05-30T01:26:30 |
https://github.com/huggingface/datasets/issues/7588
| null | 7,588 | false |
[
"Could you please run the following code snippet in your environment and share the exact output? This will help check for any compatibility issues within the env itself. \n\n```\nimport datasets\nimport huggingface_hub\nimport fsspec\n\nprint(\"datasets version:\", datasets.__version__)\nprint(\"huggingface_hub version:\", huggingface_hub.__version__)\nprint(\"fsspec version:\", fsspec.__version__)\n```",
"```bash\ndatasets version: 2.14.4\nhuggingface_hub version: 0.31.4\nfsspec version: 2025.3.2\n```",
"Version 2.14.4 is not the latest version available, in fact it is from August 08, 2023 (you can check here: https://pypi.org/project/datasets/#history)\n\nUse pip install datasets==3.6.0 to install a more recent version (from May 7, 2025)\n\nI also had the same problem with Colab, after updating to the latest version it was solved.\n\nI hope it helps",
"thank you @CleitonOERocha. it sure did help.\n\nupdating `datasets` to v3.6.0 and keeping `fsspec` on v2025.3.2 eliminates the issue.",
"Very helpful, thank you!"
] |
3,091,834,987 |
load_dataset splits typing
|
closed
|
close https://github.com/huggingface/datasets/issues/7583
| 2025-05-26T18:28:40 | 2025-05-26T18:31:10 | 2025-05-26T18:29:57 |
https://github.com/huggingface/datasets/pull/7587
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7587",
"html_url": "https://github.com/huggingface/datasets/pull/7587",
"diff_url": "https://github.com/huggingface/datasets/pull/7587.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7587.patch",
"merged_at": "2025-05-26T18:29:57"
}
| 7,587 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7587). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,091,320,431 |
help is appreciated
|
open
|
### Feature request
https://github.com/rajasekarnp1/neural-audio-upscaler/tree/main
### Motivation
ai model develpment and audio
### Your contribution
ai model develpment and audio
| 2025-05-26T14:00:42 | 2025-05-26T18:21:57 | null |
https://github.com/huggingface/datasets/issues/7586
| null | 7,586 | false |
[
"how is this related to this repository ?"
] |
3,091,227,921 |
Avoid multiple default config names
|
closed
|
Fix duplicating default config names.
Currently, when calling `push_to_hub(set_default=True` with 2 different config names, both are set as default.
Moreover, this will generate an error next time we try to push another default config name, raised by `MetadataConfigs.get_default_config_name`:
https://github.com/huggingface/datasets/blob/da1db8a5b89fc0badaa0f571b36e122e52ae8c61/src/datasets/arrow_dataset.py#L5757
https://github.com/huggingface/datasets/blob/da1db8a5b89fc0badaa0f571b36e122e52ae8c61/src/datasets/utils/metadata.py#L186-L188
| 2025-05-26T13:27:59 | 2025-06-05T12:41:54 | 2025-06-05T12:41:52 |
https://github.com/huggingface/datasets/pull/7585
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7585",
"html_url": "https://github.com/huggingface/datasets/pull/7585",
"diff_url": "https://github.com/huggingface/datasets/pull/7585.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7585.patch",
"merged_at": "2025-06-05T12:41:52"
}
| 7,585 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7585). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,090,255,023 |
Add LMDB format support
|
open
|
### Feature request
Add LMDB format support for large memory-mapping files
### Motivation
Add LMDB format support for large memory-mapping files
### Your contribution
I'm trying to add it
| 2025-05-26T07:10:13 | 2025-05-26T18:23:37 | null |
https://github.com/huggingface/datasets/issues/7584
| null | 7,584 | false |
[
"Hi ! Can you explain what's your use case ? Is it about converting LMDB to Dataset objects (i.e. converting to Arrow) ?"
] |
3,088,987,757 |
load_dataset type stubs reject List[str] for split parameter, but runtime supports it
|
closed
|
### Describe the bug
The [load_dataset](https://huggingface.co/docs/datasets/v3.6.0/en/package_reference/loading_methods#datasets.load_dataset) method accepts a `List[str]` as the split parameter at runtime, however, the current type stubs restrict the split parameter to `Union[str, Split, None]`. This causes type checkers like Pylance to raise `reportArgumentType` errors when passing a list of strings, even though it works as intended at runtime.
### Steps to reproduce the bug
1. Use load_dataset with multiple splits e.g.:
```
from datasets import load_dataset
ds_train, ds_val, ds_test = load_dataset(
"Silly-Machine/TuPyE-Dataset",
"binary",
split=["train[:75%]", "train[75%:]", "test"]
)
```
2. Observe that code executes correctly at runtime and Pylance raises `Argument of type "List[str]" cannot be assigned to parameter "split" of type "str | Split | None"`
### Expected behavior
The type stubs for [load_dataset](https://huggingface.co/docs/datasets/v3.6.0/en/package_reference/loading_methods#datasets.load_dataset) should accept `Union[str, Split, List[str], None]` or more specific overloads for the split parameter to correctly represent runtime behavior.
### Environment info
- `datasets` version: 3.6.0
- Platform: Linux-5.15.167.4-microsoft-standard-WSL2-x86_64-with-glibc2.39
- Python version: 3.12.7
- `huggingface_hub` version: 0.32.0
- PyArrow version: 20.0.0
- Pandas version: 2.2.3
- `fsspec` version: 2025.3.0
| 2025-05-25T02:33:18 | 2025-05-26T18:29:58 | 2025-05-26T18:29:58 |
https://github.com/huggingface/datasets/issues/7583
| null | 7,583 | false |
[] |
3,083,515,643 |
fix: Add embed_storage in Pdf feature
|
closed
|
Add missing `embed_storage` method in Pdf feature (Same as in Audio and Image)
| 2025-05-22T14:06:29 | 2025-05-22T14:17:38 | 2025-05-22T14:17:36 |
https://github.com/huggingface/datasets/pull/7582
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7582",
"html_url": "https://github.com/huggingface/datasets/pull/7582",
"diff_url": "https://github.com/huggingface/datasets/pull/7582.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7582.patch",
"merged_at": "2025-05-22T14:17:36"
}
| 7,582 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7582). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,083,080,413 |
Add missing property on `RepeatExamplesIterable`
|
closed
|
Fixes #7561
| 2025-05-22T11:41:07 | 2025-06-05T12:41:30 | 2025-06-05T12:41:29 |
https://github.com/huggingface/datasets/pull/7581
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7581",
"html_url": "https://github.com/huggingface/datasets/pull/7581",
"diff_url": "https://github.com/huggingface/datasets/pull/7581.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7581.patch",
"merged_at": "2025-06-05T12:41:29"
}
| 7,581 | true |
[] |
3,082,993,027 |
Requesting a specific split (eg: test) still downloads all (train, test, val) data when streaming=False.
|
open
|
### Describe the bug
When using load_dataset() from the datasets library (in load.py), specifying a particular split (e.g., split="train") still results in downloading data for all splits when streaming=False. This happens during the builder_instance.download_and_prepare() call.
This behavior leads to unnecessary bandwidth usage and longer download times, especially for large datasets, even if the user only intends to use a single split.
### Steps to reproduce the bug
dataset_name = "skbose/indian-english-nptel-v0"
dataset = load_dataset(dataset_name, token=hf_token, split="test")
### Expected behavior
Optimize the download logic so that only the required split is downloaded when streaming=False when a specific split is provided.
### Environment info
Dataset: skbose/indian-english-nptel-v0
Platform: M1 Apple Silicon
Python verison: 3.12.9
datasets>=3.5.0
| 2025-05-22T11:08:16 | 2025-05-26T18:40:31 | null |
https://github.com/huggingface/datasets/issues/7580
| null | 7,580 | false |
[
"Hi ! There was a PR open to improve this: https://github.com/huggingface/datasets/pull/6832 \nbut it hasn't been continued so far.\n\nIt would be a cool improvement though !"
] |
3,081,849,022 |
Fix typos in PDF and Video documentation
|
closed
| null | 2025-05-22T02:27:40 | 2025-05-22T12:53:49 | 2025-05-22T12:53:47 |
https://github.com/huggingface/datasets/pull/7579
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7579",
"html_url": "https://github.com/huggingface/datasets/pull/7579",
"diff_url": "https://github.com/huggingface/datasets/pull/7579.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7579.patch",
"merged_at": "2025-05-22T12:53:47"
}
| 7,579 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7579). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,080,833,740 |
arrow_schema is not compatible with list
|
closed
|
### Describe the bug
```
import datasets
f = datasets.Features({'x': list[datasets.Value(dtype='int32')]})
f.arrow_schema
Traceback (most recent call last):
File "datasets/features/features.py", line 1826, in arrow_schema
return pa.schema(self.type).with_metadata({"huggingface": json.dumps(hf_metadata)})
^^^^^^^^^
File "datasets/features/features.py", line 1815, in type
return get_nested_type(self)
^^^^^^^^^^^^^^^^^^^^^
File "datasets/features/features.py", line 1252, in get_nested_type
return pa.struct(
^^^^^^^^^^
File "pyarrow/types.pxi", line 5406, in pyarrow.lib.struct
File "pyarrow/types.pxi", line 3890, in pyarrow.lib.field
File "pyarrow/types.pxi", line 5918, in pyarrow.lib.ensure_type
TypeError: DataType expected, got <class 'list'>
```
The following works
```
f = datasets.Features({'x': datasets.LargeList(datasets.Value(dtype='int32'))})
```
### Expected behavior
according to https://github.com/huggingface/datasets/blob/458f45a22c3cc9aea5f442f6f519333dcfeae9b9/src/datasets/features/features.py#L1765 python list should be a valid type specification for features
### Environment info
- `datasets` version: 3.5.1
- Platform: macOS-15.5-arm64-arm-64bit
- Python version: 3.12.9
- `huggingface_hub` version: 0.30.2
- PyArrow version: 19.0.1
- Pandas version: 2.2.3
- `fsspec` version: 2024.12.0
| 2025-05-21T16:37:01 | 2025-05-26T18:49:51 | 2025-05-26T18:32:55 |
https://github.com/huggingface/datasets/issues/7577
| null | 7,577 | false |
[
"Thanks for reporting, I'll look into it",
"Actually it looks like you just forgot parenthesis:\n\n```diff\n- f = datasets.Features({'x': list[datasets.Value(dtype='int32')]})\n+ f = datasets.Features({'x': list([datasets.Value(dtype='int32')])})\n```\n\nor simply using the `[ ]` syntax:\n\n```python\nf = datasets.Features({'x':[datasets.Value(dtype='int32')]})\n```\n\nI'm closing this issue if you don't mind",
"Ah is that what the syntax is? I don't think I was able to find an actual example of it so I assumed it was in the same way that you specify types eg. `list[int]`. This is good to know, thanks."
] |
3,080,450,538 |
Fix regex library warnings
|
closed
|
# PR Summary
This small PR resolves the regex library warnings showing starting Python3.11:
```python
DeprecationWarning: 'count' is passed as positional argument
```
| 2025-05-21T14:31:58 | 2025-06-05T13:35:16 | 2025-06-05T12:37:55 |
https://github.com/huggingface/datasets/pull/7576
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7576",
"html_url": "https://github.com/huggingface/datasets/pull/7576",
"diff_url": "https://github.com/huggingface/datasets/pull/7576.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7576.patch",
"merged_at": "2025-06-05T12:37:55"
}
| 7,576 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7576). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,080,228,718 |
[MINOR:TYPO] Update save_to_disk docstring
|
closed
|
r/hub/filesystem in save_to_disk
| 2025-05-21T13:22:24 | 2025-06-05T12:39:13 | 2025-06-05T12:39:13 |
https://github.com/huggingface/datasets/pull/7575
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7575",
"html_url": "https://github.com/huggingface/datasets/pull/7575",
"diff_url": "https://github.com/huggingface/datasets/pull/7575.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7575.patch",
"merged_at": "2025-06-05T12:39:13"
}
| 7,575 | true |
[] |
3,079,641,072 |
Missing multilingual directions in IWSLT2017 dataset's processing script
|
open
|
### Describe the bug
Hi,
Upon using `iwslt2017.py` in `IWSLT/iwslt2017` on the Hub for loading the datasets, I am unable to obtain the datasets for the language pairs `de-it`, `de-ro`, `de-nl`, `it-de`, `nl-de`, and `ro-de` using it. These 6 pairs do not show up when using `get_dataset_config_names()` to obtain the list of all the configs present in `IWSLT/iwslt2017`. This should not be the case since as mentioned in their original paper (please see https://aclanthology.org/2017.iwslt-1.1.pdf), the authors specify that "_this year we proposed the multilingual translation between any pair of languages from {Dutch, English, German, Italian, Romanian}..._" and because these datasets are indeed present in `data/2017-01-trnmted/texts/DeEnItNlRo/DeEnItNlRo/DeEnItNlRo-DeEnItNlRo.zip`.
Best Regards,
Anand
### Steps to reproduce the bug
Check the output of `get_dataset_config_names("IWSLT/iwslt2017", trust_remote_code=True)`: only 24 language pairs are present and the following 6 config names are absent: `iwslt2017-de-it`, `iwslt2017-de-ro`, `iwslt2017-de-nl`, `iwslt2017-it-de`, `iwslt2017-nl-de`, and `iwslt2017-ro-de`.
### Expected behavior
The aforementioned 6 language pairs should also be present and hence, all these 6 language pairs' IWSLT2017 datasets must also be available for further use.
I would suggest removing `de` from the `BI_LANGUAGES` list and moving it over to the `MULTI_LANGUAGES` list instead in `iwslt2017.py` to account for all the 6 missing language pairs (the same `de-en` dataset is present in both `data/2017-01-trnmted/texts/DeEnItNlRo/DeEnItNlRo/DeEnItNlRo-DeEnItNlRo.zip` and `data/2017-01-trnted/texts/de/en/de-en.zip` but the `de-ro`, `de-nl`, `it-de`, `nl-de`, and `ro-de` datasets are only present in `data/2017-01-trnmted/texts/DeEnItNlRo/DeEnItNlRo/DeEnItNlRo-DeEnItNlRo.zip`: so, its unclear why the following comment: _`# XXX: Artificially removed DE from here, as it also exists within bilingual data`_ has been added as `L71` in `iwslt2017.py`). The `README.md` file in `IWSLT/iwslt2017`must then be re-created using `datasets-cli test path/to/iwslt2017.py --save_info --all_configs` to pass all split size verification checks for the 6 new language pairs which were previously non-existent.
### Environment info
- `datasets` version: 3.5.0
- Platform: Linux-6.8.0-56-generic-x86_64-with-glibc2.39
- Python version: 3.12.3
- `huggingface_hub` version: 0.30.1
- PyArrow version: 19.0.1
- Pandas version: 2.2.3
- `fsspec` version: 2024.12.0
| 2025-05-21T09:53:17 | 2025-05-26T18:36:38 | null |
https://github.com/huggingface/datasets/issues/7574
| null | 7,574 | false |
[
"I have opened 2 PRs on the Hub: `https://huggingface.co/datasets/IWSLT/iwslt2017/discussions/7` and `https://huggingface.co/datasets/IWSLT/iwslt2017/discussions/8` to resolve this issue",
"cool ! I pinged the owners of the dataset on HF to merge your PRs :)"
] |
3,076,415,382 |
No Samsum dataset
|
closed
|
### Describe the bug
https://huggingface.co/datasets/Samsung/samsum dataset not found error 404
Originated from https://github.com/meta-llama/llama-cookbook/issues/948
### Steps to reproduce the bug
go to website https://huggingface.co/datasets/Samsung/samsum
see the error
also downloading it with python throws
```
Couldn't find 'Samsung/samsum' on the Hugging Face Hub either: FileNotFoundError: Samsung/samsum@f00baf5a7d4abfec6820415493bcb52c587788e6/samsum.py (repository not found)
```
### Expected behavior
Dataset exists
### Environment info
```
- `datasets` version: 3.2.0
- Platform: macOS-15.4.1-arm64-arm-64bit
- Python version: 3.12.2
- `huggingface_hub` version: 0.26.5
- PyArrow version: 16.1.0
- Pandas version: 2.2.3
- `fsspec` version: 2024.9.0
```
| 2025-05-20T09:54:35 | 2025-07-21T18:34:34 | 2025-06-18T12:52:23 |
https://github.com/huggingface/datasets/issues/7573
| null | 7,573 | false |
[
"According to the following https://huggingface.co/posts/seawolf2357/424129432408590, as of now the dataset seems to be inaccessible.\n\n@IgorKasianenko, would https://huggingface.co/datasets/knkarthick/samsum suffice for your purpose?\n",
"Thanks @SP1029 for the update!\nThat will work for now, using it as replacement. Is there a officially recommended way to maintain the CC licensed dataset under the organization account? \nFeel free to close this issue",
"> Is there an officially recommended way to maintain a CC-licensed dataset under an organizational account?\n\n@IgorKasianenko, apologies, this is not my area of expertise.\n\n> Please feel free to close this issue.\n\nI have limited access and may not be able to do that. Since you opened it, you would be able to close it.",
"dataset_samsum = load_dataset(\"knkarthick/samsum\")\n\nis working"
] |
3,074,529,251 |
Fixed typos
|
closed
|
More info: [comment](https://github.com/huggingface/datasets/pull/7564#issuecomment-2863391781).
| 2025-05-19T17:16:59 | 2025-06-05T12:25:42 | 2025-06-05T12:25:41 |
https://github.com/huggingface/datasets/pull/7572
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7572",
"html_url": "https://github.com/huggingface/datasets/pull/7572",
"diff_url": "https://github.com/huggingface/datasets/pull/7572.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7572.patch",
"merged_at": "2025-06-05T12:25:41"
}
| 7,572 | true |
[
"@lhoestq, mentioning in case you haven't seen this PR. The contribution is very small and easy to check :)"
] |
3,074,116,942 |
fix string_to_dict test
|
closed
| null | 2025-05-19T14:49:23 | 2025-05-19T14:52:24 | 2025-05-19T14:49:28 |
https://github.com/huggingface/datasets/pull/7571
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7571",
"html_url": "https://github.com/huggingface/datasets/pull/7571",
"diff_url": "https://github.com/huggingface/datasets/pull/7571.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7571.patch",
"merged_at": "2025-05-19T14:49:28"
}
| 7,571 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7571). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,065,966,529 |
Dataset lib seems to broke after fssec lib update
|
closed
|
### Describe the bug
I am facing an issue since today where HF's dataset is acting weird and in some instances failure to recognise a valid dataset entirely, I think it is happening due to recent change in `fsspec` lib as using this command fixed it for me in one-time: `!pip install -U datasets huggingface_hub fsspec`
### Steps to reproduce the bug
from datasets import load_dataset
def download_hf():
dataset_name = input("Enter the dataset name: ")
subset_name = input("Enter subset name: ")
ds = load_dataset(dataset_name, name=subset_name)
for split in ds:
ds[split].to_pandas().to_csv(f"{subset_name}.csv", index=False)
download_hf()
### Expected behavior
```
Downloading readme: 100%
1.55k/1.55k [00:00<00:00, 121kB/s]
Downloading data files: 100%
1/1 [00:00<00:00, 2.06it/s]
Downloading data: 0%| | 0.00/54.2k [00:00<?, ?B/s]
Downloading data: 100%|██████████| 54.2k/54.2k [00:00<00:00, 121kB/s]
Extracting data files: 100%
1/1 [00:00<00:00, 35.17it/s]
Generating test split:
140/0 [00:00<00:00, 2628.62 examples/s]
---------------------------------------------------------------------------
NotImplementedError Traceback (most recent call last)
[<ipython-input-2-12ab305b0e77>](https://localhost:8080/#) in <cell line: 0>()
8 ds[split].to_pandas().to_csv(f"{subset_name}.csv", index=False)
9
---> 10 download_hf()
2 frames
[/usr/local/lib/python3.11/dist-packages/datasets/builder.py](https://localhost:8080/#) in as_dataset(self, split, run_post_process, verification_mode, ignore_verifications, in_memory)
1171 is_local = not is_remote_filesystem(self._fs)
1172 if not is_local:
-> 1173 raise NotImplementedError(f"Loading a dataset cached in a {type(self._fs).__name__} is not supported.")
1174 if not os.path.exists(self._output_dir):
1175 raise FileNotFoundError(
NotImplementedError: Loading a dataset cached in a LocalFileSystem is not supported.
```
OR
```
Traceback (most recent call last):
File "e:\Fuck\download-data\mcq_dataset.py", line 10, in <module>
download_hf()
File "e:\Fuck\download-data\mcq_dataset.py", line 6, in download_hf
ds = load_dataset(dataset_name, name=subset_name)
File "C:\Users\DELL\AppData\Local\Programs\Python\Python310\lib\site-packages\datasets\load.py", line 2606, in load_dataset
builder_instance = load_dataset_builder(
File "C:\Users\DELL\AppData\Local\Programs\Python\Python310\lib\site-packages\datasets\load.py", line 2277, in load_dataset_builder
dataset_module = dataset_module_factory(
File "C:\Users\DELL\AppData\Local\Programs\Python\Python310\lib\site-packages\datasets\load.py", line 1917, in dataset_module_factory
raise e1 from None
File "C:\Users\DELL\AppData\Local\Programs\Python\Python310\lib\site-packages\datasets\load.py", line 1867, in dataset_module_factory
raise DatasetNotFoundError(f"Dataset '{path}' doesn't exist on the Hub or cannot be accessed.") from e
datasets.exceptions.DatasetNotFoundError: Dataset 'dataset repo_id' doesn't exist on the Hub or cannot be accessed.
```
### Environment info
colab and 3.10 local system
| 2025-05-15T11:45:06 | 2025-06-13T00:44:27 | 2025-06-13T00:44:27 |
https://github.com/huggingface/datasets/issues/7570
| null | 7,570 | false |
[
"Hi, can you try updating `datasets` ? Colab still installs `datasets` 2.x by default, instead of 3.x\n\nIt would be cool to also report this to google colab, they have a GitHub repo for this IIRC",
"@lhoestq I have updated it to `datasets==3.6.0` and now there's an entirely different issue on colab while locally its fine. \n\n```\n/usr/local/lib/python3.11/dist-packages/huggingface_hub/utils/_auth.py:94: UserWarning: \nThe secret `HF_TOKEN` does not exist in your Colab secrets.\nTo authenticate with the Hugging Face Hub, create a token in your settings tab (https://huggingface.co/settings/tokens), set it as secret in your Google Colab and restart your session.\nYou will be able to reuse this secret in all of your notebooks.\nPlease note that authentication is recommended but still optional to access public models or datasets.\n warnings.warn(\nREADME.md: 100%\n 2.88k/2.88k [00:00<00:00, 166kB/s]\nsuno.jsonl.zst: 100%\n 221M/221M [00:05<00:00, 48.6MB/s]\nGenerating train split: \n 18633/0 [00:01<00:00, 13018.92 examples/s]\n---------------------------------------------------------------------------\nTypeError Traceback (most recent call last)\n[/usr/local/lib/python3.11/dist-packages/datasets/builder.py](https://localhost:8080/#) in _prepare_split_single(self, gen_kwargs, fpath, file_format, max_shard_size, job_id)\n 1870 try:\n-> 1871 writer.write_table(table)\n 1872 except CastError as cast_error:\n\n17 frames\nTypeError: Couldn't cast array of type\nstruct<id: string, type: string, infill: bool, source: string, continue_at: double, infill_dur_s: double, infill_end_s: double, infill_start_s: double, include_future_s: double, include_history_s: double, infill_context_end_s: double, infill_context_start_s: int64>\nto\n{'id': Value(dtype='string', id=None), 'type': Value(dtype='string', id=None), 'infill': Value(dtype='bool', id=None), 'source': Value(dtype='string', id=None), 'continue_at': Value(dtype='float64', id=None), 'include_history_s': Value(dtype='float64', id=None)}\n\nThe above exception was the direct cause of the following exception:\n\nDatasetGenerationError Traceback (most recent call last)\n[/usr/local/lib/python3.11/dist-packages/datasets/builder.py](https://localhost:8080/#) in _prepare_split_single(self, gen_kwargs, fpath, file_format, max_shard_size, job_id)\n 1896 if isinstance(e, DatasetGenerationError):\n 1897 raise\n-> 1898 raise DatasetGenerationError(\"An error occurred while generating the dataset\") from e\n 1899 \n 1900 yield job_id, True, (total_num_examples, total_num_bytes, writer._features, num_shards, shard_lengths)\n\nDatasetGenerationError: An error occurred while generating the dataset\n```",
"@lhoestq opps sorry the dataset was in .zst which was causing this error rather than being a datasets library fault. After upgrading dataset version Colab is working fine. "
] |
3,061,234,054 |
Dataset creation is broken if nesting a dict inside a dict inside a list
|
open
|
### Describe the bug
Hey,
I noticed that the creation of datasets with `Dataset.from_generator` is broken if dicts and lists are nested in a certain way and a schema is being passed. See below for details.
Best,
Tim
### Steps to reproduce the bug
Runing this code:
```python
from datasets import Dataset, Features, Sequence, Value
def generator():
yield {
"a": [{"b": {"c": 0}}],
}
features = Features(
{
"a": Sequence(
feature={
"b": {
"c": Value("int32"),
},
},
length=1,
)
}
)
dataset = Dataset.from_generator(generator, features=features)
```
leads to
```
Generating train split: 1 examples [00:00, 540.85 examples/s]
Traceback (most recent call last):
File "/home/user/miniconda3/envs/test/lib/python3.11/site-packages/datasets/builder.py", line 1635, in _prepare_split_single
num_examples, num_bytes = writer.finalize()
^^^^^^^^^^^^^^^^^
File "/home/user/miniconda3/envs/test/lib/python3.11/site-packages/datasets/arrow_writer.py", line 657, in finalize
self.write_examples_on_file()
File "/home/user/miniconda3/envs/test/lib/python3.11/site-packages/datasets/arrow_writer.py", line 510, in write_examples_on_file
self.write_batch(batch_examples=batch_examples)
File "/home/user/miniconda3/envs/test/lib/python3.11/site-packages/datasets/arrow_writer.py", line 629, in write_batch
pa_table = pa.Table.from_arrays(arrays, schema=schema)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "pyarrow/table.pxi", line 4851, in pyarrow.lib.Table.from_arrays
File "pyarrow/table.pxi", line 1608, in pyarrow.lib._sanitize_arrays
File "pyarrow/array.pxi", line 399, in pyarrow.lib.asarray
File "pyarrow/array.pxi", line 1004, in pyarrow.lib.Array.cast
File "/home/user/miniconda3/envs/test/lib/python3.11/site-packages/pyarrow/compute.py", line 405, in cast
return call_function("cast", [arr], options, memory_pool)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "pyarrow/_compute.pyx", line 598, in pyarrow._compute.call_function
File "pyarrow/_compute.pyx", line 393, in pyarrow._compute.Function.call
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
pyarrow.lib.ArrowNotImplementedError: Unsupported cast from fixed_size_list<item: struct<c: int32>>[1] to struct using function cast_struct
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/home/user/test/tools/hf_test2.py", line 23, in <module>
dataset = Dataset.from_generator(generator, features=features)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/miniconda3/envs/test/lib/python3.11/site-packages/datasets/arrow_dataset.py", line 1114, in from_generator
).read()
^^^^^^
File "/home/user/miniconda3/envs/test/lib/python3.11/site-packages/datasets/io/generator.py", line 49, in read
self.builder.download_and_prepare(
File "/home/user/miniconda3/envs/test/lib/python3.11/site-packages/datasets/builder.py", line 925, in download_and_prepare
self._download_and_prepare(
File "/home/user/miniconda3/envs/test/lib/python3.11/site-packages/datasets/builder.py", line 1649, in _download_and_prepare
super()._download_and_prepare(
File "/home/user/miniconda3/envs/test/lib/python3.11/site-packages/datasets/builder.py", line 1001, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/user/miniconda3/envs/test/lib/python3.11/site-packages/datasets/builder.py", line 1487, in _prepare_split
for job_id, done, content in self._prepare_split_single(
File "/home/user/miniconda3/envs/test/lib/python3.11/site-packages/datasets/builder.py", line 1644, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset
Process finished with exit code 1
```
### Expected behavior
I expected this code not to lead to an error.
I have done some digging and figured out that the problem seems to be the `get_nested_type` function in `features.py`, which, for whatever reason, flips Sequences and dicts whenever it encounters a dict inside of a sequence. This seems to be necessary, as disabling that flip leads to another error. However, by keeping that flip enabled for the highest level and disabling it for all subsequent levels, I was able to work around this problem. Specifically, by patching `get_nested_type` as follows, it works on the given example (emphasis on the `level` parameter I added):
```python
def get_nested_type(schema: FeatureType, level=0) -> pa.DataType:
"""
get_nested_type() converts a datasets.FeatureType into a pyarrow.DataType, and acts as the inverse of
generate_from_arrow_type().
It performs double-duty as the implementation of Features.type and handles the conversion of
datasets.Feature->pa.struct
"""
# Nested structures: we allow dict, list/tuples, sequences
if isinstance(schema, Features):
return pa.struct(
{key: get_nested_type(schema[key], level = level + 1) for key in schema}
) # Features is subclass of dict, and dict order is deterministic since Python 3.6
elif isinstance(schema, dict):
return pa.struct(
{key: get_nested_type(schema[key], level = level + 1) for key in schema}
) # however don't sort on struct types since the order matters
elif isinstance(schema, (list, tuple)):
if len(schema) != 1:
raise ValueError("When defining list feature, you should just provide one example of the inner type")
value_type = get_nested_type(schema[0], level = level + 1)
return pa.list_(value_type)
elif isinstance(schema, LargeList):
value_type = get_nested_type(schema.feature, level = level + 1)
return pa.large_list(value_type)
elif isinstance(schema, Sequence):
value_type = get_nested_type(schema.feature, level = level + 1)
# We allow to reverse list of dict => dict of list for compatibility with tfds
if isinstance(schema.feature, dict) and level == 1:
data_type = pa.struct({f.name: pa.list_(f.type, schema.length) for f in value_type})
else:
data_type = pa.list_(value_type, schema.length)
return data_type
# Other objects are callable which returns their data type (ClassLabel, Array2D, Translation, Arrow datatype creation methods)
return schema()
```
I have honestly no idea what I am doing here, so this might produce other issues for different inputs.
### Environment info
- `datasets` version: 3.6.0
- Platform: Linux-6.8.0-59-generic-x86_64-with-glibc2.35
- Python version: 3.11.11
- `huggingface_hub` version: 0.30.2
- PyArrow version: 19.0.1
- Pandas version: 2.2.3
- `fsspec` version: 2024.12.0
Also tested it with 3.5.0, same result.
| 2025-05-13T21:06:45 | 2025-05-20T19:25:15 | null |
https://github.com/huggingface/datasets/issues/7569
| null | 7,569 | false |
[
"Hi ! That's because Séquence is a type that comes from tensorflow datasets and inverts lists and focus when doing Séquence(dict).\n\nInstead you should use a list. In your case\n```python\nfeatures = Features({\n \"a\": [{\"b\": {\"c\": Value(\"string\")}}]\n})\n```",
"Hi,\n\nThanks for the swift reply! Could you quickly clarify a couple of points?\n\n1. Is there any benefit in using Sequence over normal lists? Especially for longer lists (in my case, up to 256 entries)\n2. When exactly can I use Sequence? If there is a maximum of one level of dictionaries inside, then it's always fine?\n3. When creating the data in the generator, do I need to swap lists and dicts manually, or does that happen automatically?\n\nAlso, the documentation does not seem to mention this limitation of the Sequence type anywhere and encourages users to use it [here](https://huggingface.co/docs/datasets/en/about_dataset_features). In fact, I did not even know that just using a Python list was an option. Maybe the documentation can be improved to mention the limitations of Sequence and highlight that lists can be used instead.\n\nThanks a lot in advance!\n\nBest,\nTim"
] |
3,060,515,257 |
`IterableDatasetDict.map()` call removes `column_names` (in fact info.features)
|
open
|
When calling `IterableDatasetDict.map()`, each split’s `IterableDataset.map()` is invoked without a `features` argument. While omitting the argument isn’t itself incorrect, the implementation then sets `info.features = features`, which destroys the original `features` content. Since `IterableDataset.column_names` relies on `info.features`, it ends up broken (`None`).
**Reproduction**
1. Define an IterableDatasetDict with a non-None features schema.
2. my_iterable_dataset_dict contains "text" column.
3. Call:
```Python
new_dict = my_iterable_dataset_dict.map(
function=my_fn,
with_indices=False,
batched=True,
batch_size=16,
)
```
4. Observe
```Python
new_dict["train"].info.features # {'text': Value(dtype='string', id=None)}
new_dict["train"].column_names # ['text']
```
5. Call:
```Python
new_dict = my_iterable_dataset_dict.map(
function=my_fn,
with_indices=False,
batched=True,
batch_size=16,
remove_columns=["foo"]
)
```
6. Observe:
```Python
new_dict["train"].info.features # → None
new_dict["train"].column_names # → None
```
5. Internally, in dataset_dict.py this loop omits features ([code](https://github.com/huggingface/datasets/blob/b9efdc64c3bfb8f21f8a4a22b21bddd31ecd5a31/src/datasets/dataset_dict.py#L2047C5-L2056C14)):
```Python
for split, dataset in self.items():
dataset_dict[split] = dataset.map(
function=function,
with_indices=with_indices,
input_columns=input_columns,
batched=batched,
batch_size=batch_size,
drop_last_batch=drop_last_batch,
remove_columns=remove_columns,
fn_kwargs=fn_kwargs,
# features omitted → defaults to None
)
```
7. Then inside IterableDataset.map() ([code](https://github.com/huggingface/datasets/blob/b9efdc64c3bfb8f21f8a4a22b21bddd31ecd5a31/src/datasets/iterable_dataset.py#L2619C1-L2622C37)) correct `info.features` is replaced by features which is None:
```Python
info = self.info.copy()
info.features = features # features is None here
return IterableDataset(..., info=info, ...)
```
**Suggestion**
It looks like this replacement was added intentionally but maybe should be done only if `features` is `not None`.
**Workarround:**
`SFTTrainer` calls `dataset.map()` several times and then fails on `NoneType` when iterating `dataset.column_names`.
I decided to write this patch - works form me.
```python
def patch_iterable_dataset_map():
_orig_map = IterableDataset.map
def _patched_map(self, *args, **kwargs):
if "features" not in kwargs or kwargs["features"] is None:
kwargs["features"] = self.info.features
return _orig_map(self, *args, **kwargs)
IterableDataset.map = _patched_map
```
| 2025-05-13T15:45:42 | 2025-06-30T09:33:47 | null |
https://github.com/huggingface/datasets/issues/7568
| null | 7,568 | false |
[
"Hi ! IterableDataset doesn't know what's the output of the function you pass to map(), so it's not possible to know in advance the features of the output dataset.\n\nThere is a workaround though: either do `ds = ds.map(..., features=features)`, or you can do `ds = ds._resolve_features()` which iterates on the first rows to infer the dataset features.",
"Thank you. I understand that “IterableDataset doesn't know what's the output of the function”—that’s true, but:\n\nUnfortunately, the workaround you proposed **doesn’t solve** the problem. `ds.map()` is called multiple times by third-party code (i.e. `SFTTrainer`). To apply your approach, I would have to modify external library code. That’s why I decided to patch the _class_ rather than update `dataset` _objects_ (in fact, updating the object after `map()` was my initial approach, but then I realized I’m not the only one mapping an already-mapped dataset.)\n\nAs a user, I expected that after mapping I would get a new dataset with the correct column names. If, for some reason, that can’t be the default behavior, I would expect an argument—i.e. `auto_resolve_features: bool = False` — to control how my dataset is mapped if following mapping operation are called.\n\nIt’s also problematic that `column_names` are tied to `features`, which is even more confusing and forces you to inspect the source code to understand what’s going on.\n\n**New version of workaround:**\n```python\ndef patch_iterable_dataset_map():\n _orig_map = IterableDataset.map\n\n def _patched_map(self, *args, **kwargs):\n ds = _orig_map(self, *args, **kwargs)\n return ds._resolve_features()\n\n IterableDataset.map = _patched_map\n```",
"I see, maybe `.resolve_features()` should be called by default in this case in the SFTTrainer ? (or pass `features=` if the data processing always output the same features)\n\nWe can even support a new parameter `features=\"infer\"` if it would be comfortable to not use internal methods in SFTTrainer",
"I think most straightforward solution would be to reinitialize `features` from data after mapping if `feature` argument is not passed. I hink it is more intuitive behavior than just cleaning features. There is also problem in usage `.resolve_features()` in this context. I observed that it leads to `_head()` method execution and it then causes that 5 batches from dataset are iterated (`_head()` defaults to 5 batches). \nI'm not sure how it influences whole process. Are those 5 batches (in my case it's 5000 rows) used only to find `features`. Does final training/eval process \"see\" this items? How it affects IterableDataset state (current position)?",
"I checked the source code and while it indeed iterates on the first 5 rows. As a normal iteration, it does record the state in case you call `.state_dict()`, but it doesn't change the starting state. The starting state is always the beginning of the dataset, unless it is explicitly set with `.load_state_dict()`. To be clear, if you iterate on the dataset after `._resolve_features()`, it will start from the beginning of the dataset (or from a state you manually pass using `.load_state_dict()`)",
"Hi!\nI’ve opened a PR #7658 to address this issue.\n\nThe fix ensures that info.features is only updated if features is not None, preventing accidental loss of schema and column_names.\nPlease let me know if you see any edge cases or have additional concerns!\nAlso, if a test is needed for this case, happy to discuss—the fix is small, but I can add one if the maintainers prefer.\n\nThanks everyone for the clear diagnosis and suggestions in this thread!"
] |
3,058,308,538 |
interleave_datasets seed with multiple workers
|
open
|
### Describe the bug
Using interleave_datasets with multiple dataloader workers and a seed set causes the same dataset sampling order across all workers.
Should the seed be modulated with the worker id?
### Steps to reproduce the bug
See above
### Expected behavior
See above
### Environment info
- `datasets` version: 3.5.1
- Platform: macOS-15.4.1-arm64-arm-64bit
- Python version: 3.12.9
- `huggingface_hub` version: 0.30.2
- PyArrow version: 19.0.1
- Pandas version: 2.2.3
- `fsspec` version: 2024.12.0
| 2025-05-12T22:38:27 | 2025-06-29T06:53:59 | null |
https://github.com/huggingface/datasets/issues/7567
| null | 7,567 | false |
[
"Hi ! It's already the case IIRC: the effective seed looks like `seed + worker_id`. Do you have a reproducible example ?",
"here is an example with shuffle\n\n```\nimport itertools\nimport datasets\nimport multiprocessing\nimport torch.utils.data\n\n\ndef gen(shard):\n worker_info = torch.utils.data.get_worker_info()\n for i in range(10):\n yield {'value': i, 'worker_id': worker_info.id}\n\n\ndef main():\n ds = datasets.IterableDataset.from_generator(gen, gen_kwargs={'shard': list(range(8))})\n ds = ds.shuffle(buffer_size=100, seed=1234)\n dataloader = torch.utils.data.DataLoader(ds, batch_size=None, num_workers=8)\n for i, ex in enumerate(itertools.islice(dataloader, 50)):\n print(i, ex)\n\n\nif __name__ == '__main__':\n multiprocessing.set_start_method('spawn')\n main()\n```\n\n```\npython test.py\n0 {'value': 8, 'worker_id': 0}\n1 {'value': 8, 'worker_id': 1}\n2 {'value': 8, 'worker_id': 2}\n3 {'value': 8, 'worker_id': 3}\n4 {'value': 8, 'worker_id': 4}\n5 {'value': 8, 'worker_id': 5}\n6 {'value': 8, 'worker_id': 6}\n7 {'value': 8, 'worker_id': 7}\n8 {'value': 9, 'worker_id': 0}\n9 {'value': 9, 'worker_id': 1}\n10 {'value': 9, 'worker_id': 2}\n11 {'value': 9, 'worker_id': 3}\n12 {'value': 9, 'worker_id': 4}\n13 {'value': 9, 'worker_id': 5}\n14 {'value': 9, 'worker_id': 6}\n15 {'value': 9, 'worker_id': 7}\n16 {'value': 5, 'worker_id': 0}\n17 {'value': 5, 'worker_id': 1}\n18 {'value': 5, 'worker_id': 2}\n19 {'value': 5, 'worker_id': 3}\n```",
"With `interleave_datasets`\n\n```\nimport itertools\nimport datasets\nimport multiprocessing\nimport torch.utils.data\n\n\ndef gen(shard, value):\n while True:\n yield {'value': value}\n\n\ndef main():\n ds = [\n datasets.IterableDataset.from_generator(gen, gen_kwargs={'shard': list(range(8)), 'value': i})\n for i in range(10)\n ]\n ds = datasets.interleave_datasets(ds, probabilities=[1 / len(ds)] * len(ds), seed=1234)\n dataloader = torch.utils.data.DataLoader(ds, batch_size=None, num_workers=8)\n for i, ex in enumerate(itertools.islice(dataloader, 50)):\n print(i, ex)\n\n\nif __name__ == '__main__':\n multiprocessing.set_start_method('spawn')\n main()\n```\n\n```\npython test.py\n0 {'value': 9}\n1 {'value': 9}\n2 {'value': 9}\n3 {'value': 9}\n4 {'value': 9}\n5 {'value': 9}\n6 {'value': 9}\n7 {'value': 9}\n8 {'value': 3}\n9 {'value': 3}\n10 {'value': 3}\n11 {'value': 3}\n12 {'value': 3}\n13 {'value': 3}\n14 {'value': 3}\n15 {'value': 3}\n16 {'value': 9}\n17 {'value': 9}\n18 {'value': 9}\n19 {'value': 9}\n20 {'value': 9}\n21 {'value': 9}\n22 {'value': 9}\n23 {'value': 9}\n```",
"Same results after updating to datasets 3.6.0.",
"Ah my bad, `shuffle()` uses a global effective seed which is something like `seed + epoch`, which is used to do the same shards shuffle in each worker so that each worker have a non-overlapping set of shards:\n\nhttps://github.com/huggingface/datasets/blob/b9efdc64c3bfb8f21f8a4a22b21bddd31ecd5a31/src/datasets/iterable_dataset.py#L2102-L2111\n\nI think we should take into account the `worker_id` in a local seed for the buffer right after this line:\n\nhttps://github.com/huggingface/datasets/blob/b9efdc64c3bfb8f21f8a4a22b21bddd31ecd5a31/src/datasets/iterable_dataset.py#L2151-L2153\n\nlike adding a new step that would propagate in the examples iterables or something like that:\n\n```python\nex_iterable = ex_iterable.shift_rngs(value=worker_id)\n```\n\nis this something you'd like to explore ? contributions on this subject are very welcome",
"Potentially, but busy. If anyone wants to take this up please feel free to, otherwise I may or may not revisit when I have free time.\n\nFor what it's worth I got around this with\n\n```\n\nclass SeedGeneratorWithWorkerIterable(iterable_dataset._BaseExamplesIterable):\n \"\"\"ExamplesIterable that seeds the rng with worker id.\"\"\"\n\n def __init__(\n self,\n ex_iterable: iterable_dataset._BaseExamplesIterable,\n generator: np.random.Generator,\n rank: int = 0,\n ):\n \"\"\"Constructor.\"\"\"\n super().__init__()\n self.ex_iterable = ex_iterable\n self.generator = generator\n self.rank = rank\n\n def _init_state_dict(self) -> dict:\n self._state_dict = self.ex_iterable._init_state_dict()\n return self._state_dict\n\n def __iter__(self):\n \"\"\"Data iterator.\"\"\"\n effective_seed = copy.deepcopy(self.generator).integers(0, 1 << 63) - self.rank\n effective_seed = (1 << 63) + effective_seed if effective_seed < 0 else effective_seed\n generator = np.random.default_rng(effective_seed)\n self.ex_iterable = self.ex_iterable.shuffle_data_sources(generator)\n if self._state_dict:\n self._state_dict = self.ex_iterable._init_state_dict()\n yield from iter(self.ex_iterable)\n\n def shuffle_data_sources(self, generator):\n \"\"\"Shuffle data sources.\"\"\"\n ex_iterable = self.ex_iterable.shuffle_data_sources(generator)\n return SeedGeneratorWithWorkerIterable(ex_iterable, generator=generator, rank=self.rank)\n\n def shard_data_sources(self, num_shards: int, index: int, contiguous=True): # noqa: FBT002\n \"\"\"Shard data sources.\"\"\"\n ex_iterable = self.ex_iterable.shard_data_sources(num_shards, index, contiguous=contiguous)\n return SeedGeneratorWithWorkerIterable(ex_iterable, generator=self.generator, rank=index)\n\n @property\n def is_typed(self):\n return self.ex_iterable.is_typed\n\n @property\n def features(self):\n return self.ex_iterable.features\n\n @property\n def num_shards(self) -> int:\n \"\"\"Number of shards.\"\"\"\n return self.ex_iterable.num_shards\n```",
"Thanks for the detailed insights!\n\nAfter reviewing the issue and the current implementation in `iterable_dataset.py`, I can confirm the cause:\n\nWhen using `interleave_datasets(..., seed=...)` with `num_workers > 1` (e.g. via `DataLoader`), the same RNG state is shared across workers — which leads to each worker producing identical sample sequences. This is because the seed is not modulated by `worker_id`, unlike the usual approach in `shuffle()` where seed is adjusted using the `epoch`.\n\nAs @lhoestq suggested, a proper fix would involve introducing something like:\n\n```python\nex_iterable = ex_iterable.shift_rngs(worker_id)\n```\n\n@jonathanasdf Also really appreciate the workaround implementation shared above — that was helpful to validate the behavior and will help shape the general solution."
] |
3,055,279,344 |
terminate called without an active exception; Aborted (core dumped)
|
open
|
### Describe the bug
I use it as in the tutorial here: https://huggingface.co/docs/datasets/stream, and it ends up with abort.
### Steps to reproduce the bug
1. `pip install datasets`
2.
```
$ cat main.py
#!/usr/bin/env python3
from datasets import load_dataset
dataset = load_dataset('HuggingFaceFW/fineweb', split='train', streaming=True)
print(next(iter(dataset)))
```
3. `chmod +x main.py`
```
$ ./main.py
README.md: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 43.1k/43.1k [00:00<00:00, 7.04MB/s]
Resolving data files: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 25868/25868 [00:05<00:00, 4859.26it/s]
Resolving data files: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 25868/25868 [00:00<00:00, 54773.56it/s]
{'text': "How AP reported in all formats from tornado-stricken regionsMarch 8, 2012\nWhen the first serious bout of tornadoes of 2012 blew through middle America in the middle of the night, they touched down in places hours from any AP bureau. Our closest video journalist was Chicago-based Robert Ray, who dropped his plans to travel to Georgia for Super Tuesday, booked several flights to the cities closest to the strikes and headed for the airport. He’d decide once there which flight to take.\nHe never got on board a plane. Instead, he ended up driving toward Harrisburg, Ill., where initial reports suggested a town was destroyed. That decision turned out to be a lucky break for the AP. Twice.\nRay was among the first journalists to arrive and he confirmed those reports -- in all formats. He shot powerful video, put victims on the phone with AP Radio and played back sound to an editor who transcribed the interviews and put the material on text wires. He then walked around the devastation with the Central Regional Desk on the line, talking to victims with the phone held so close that editors could transcribe his interviews in real time.\nRay also made a dramatic image of a young girl who found a man’s prosthetic leg in the rubble, propped it up next to her destroyed home and spray-painted an impromptu sign: “Found leg. Seriously.”\nThe following day, he was back on the road and headed for Georgia and a Super Tuesday date with Newt Gingrich’s campaign. The drive would take him through a stretch of the South that forecasters expected would suffer another wave of tornadoes.\nTo prevent running into THAT storm, Ray used his iPhone to monitor Doppler radar, zooming in on extreme cells and using Google maps to direct himself to safe routes. And then the journalist took over again.\n“When weather like that occurs, a reporter must seize the opportunity to get the news out and allow people to see, hear and read the power of nature so that they can take proper shelter,” Ray says.\nSo Ray now started to use his phone to follow the storms. He attached a small GoPro camera to his steering wheel in case a tornado dropped down in front of the car somewhere, and took video of heavy rain and hail with his iPhone. Soon, he spotted a tornado and the chase was on. He followed an unmarked emergency vehicle to Cleveland, Tenn., where he was first on the scene of the storm's aftermath.\nAgain, the tornadoes had struck in locations that were hours from the nearest AP bureau. Damage and debris, as well as a wickedly violent storm that made travel dangerous, slowed our efforts to get to the news. That wasn’t a problem in Tennessee, where our customers were well served by an all-formats report that included this text story.\n“CLEVELAND, Tenn. (AP) _ Fierce wind, hail and rain lashed Tennessee for the second time in three days, and at least 15 people were hospitalized Friday in the Chattanooga area.”\nThe byline? Robert Ray.\nFor being adept with technology, chasing after news as it literally dropped from the sky and setting a standard for all-formats reporting that put the AP ahead on the most competitive news story of the day, Ray wins this week’s $300 Best of the States prize.\n© 2013 The Associated Press. All rights reserved. Terms and conditions apply. See AP.org for details.", 'id': '<urn:uuid:d66bc6fe-8477-4adf-b430-f6a558ccc8ff>', 'dump': 'CC-MAIN-2013-20', 'url': 'http://%[email protected]/Content/Press-Release/2012/How-AP-reported-in-all-formats-from-tornado-stricken-regions', 'date': '2013-05-18T05:48:54Z', 'file_path': 's3://commoncrawl/crawl-data/CC-MAIN-2013-20/segments/1368696381249/warc/CC-MAIN-20130516092621-00000-ip-10-60-113-184.ec2.internal.warc.gz', 'language': 'en', 'language_score': 0.9721424579620361, 'token_count': 717}
terminate called without an active exception
Aborted (core dumped)
```
### Expected behavior
I'm not a proficient Python user, so it might be my own error, but even in that case, the error message should be better.
### Environment info
`Successfully installed datasets-3.6.0 dill-0.3.8 hf-xet-1.1.0 huggingface-hub-0.31.1 multiprocess-0.70.16 requests-2.32.3 xxhash-3.5.0`
```
$ cat /etc/lsb-release
DISTRIB_ID=Ubuntu
DISTRIB_RELEASE=22.04
DISTRIB_CODENAME=jammy
DISTRIB_DESCRIPTION="Ubuntu 22.04.4 LTS"
```
| 2025-05-11T23:05:54 | 2025-06-23T17:56:02 | null |
https://github.com/huggingface/datasets/issues/7566
| null | 7,566 | false |
[
"@alexey-milovidov I followed the code snippet, but am able to successfully execute without any error. Could you please verify if the error persists or there is any additional details.",
"@alexey-milovidov else if the problem does not exist please feel free to close this issue.",
"```\nmilovidov@milovidov-pc:~/work/datasets$ \n./main.py \nResolving data files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 25868/25868 [00:05<00:00, 4753.90it/s]\nResolving data files: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 25868/25868 [00:00<00:00, 238798.85it/s]\n{'text': \"How AP reported in all formats from tornado-stricken regionsMarch 8, 2012\\nWhen the first serious bout of tornadoes of 2012 blew through middle America in the middle of the night, they touched down in places hours from any AP bureau. Our closest video journalist was Chicago-based Robert Ray, who dropped his plans to travel to Georgia for Super Tuesday, booked several flights to the cities closest to the strikes and headed for the airport. He’d decide once there which flight to take.\\nHe never got on board a plane. Instead, he ended up driving toward Harrisburg, Ill., where initial reports suggested a town was destroyed. That decision turned out to be a lucky break for the AP. Twice.\\nRay was among the first journalists to arrive and he confirmed those reports -- in all formats. He shot powerful video, put victims on the phone with AP Radio and played back sound to an editor who transcribed the interviews and put the material on text wires. He then walked around the devastation with the Central Regional Desk on the line, talking to victims with the phone held so close that editors could transcribe his interviews in real time.\\nRay also made a dramatic image of a young girl who found a man’s prosthetic leg in the rubble, propped it up next to her destroyed home and spray-painted an impromptu sign: “Found leg. Seriously.”\\nThe following day, he was back on the road and headed for Georgia and a Super Tuesday date with Newt Gingrich’s campaign. The drive would take him through a stretch of the South that forecasters expected would suffer another wave of tornadoes.\\nTo prevent running into THAT storm, Ray used his iPhone to monitor Doppler radar, zooming in on extreme cells and using Google maps to direct himself to safe routes. And then the journalist took over again.\\n“When weather like that occurs, a reporter must seize the opportunity to get the news out and allow people to see, hear and read the power of nature so that they can take proper shelter,” Ray says.\\nSo Ray now started to use his phone to follow the storms. He attached a small GoPro camera to his steering wheel in case a tornado dropped down in front of the car somewhere, and took video of heavy rain and hail with his iPhone. Soon, he spotted a tornado and the chase was on. He followed an unmarked emergency vehicle to Cleveland, Tenn., where he was first on the scene of the storm's aftermath.\\nAgain, the tornadoes had struck in locations that were hours from the nearest AP bureau. Damage and debris, as well as a wickedly violent storm that made travel dangerous, slowed our efforts to get to the news. That wasn’t a problem in Tennessee, where our customers were well served by an all-formats report that included this text story.\\n“CLEVELAND, Tenn. (AP) _ Fierce wind, hail and rain lashed Tennessee for the second time in three days, and at least 15 people were hospitalized Friday in the Chattanooga area.”\\nThe byline? Robert Ray.\\nFor being adept with technology, chasing after news as it literally dropped from the sky and setting a standard for all-formats reporting that put the AP ahead on the most competitive news story of the day, Ray wins this week’s $300 Best of the States prize.\\n© 2013 The Associated Press. All rights reserved. Terms and conditions apply. See AP.org for details.\", 'id': '<urn:uuid:d66bc6fe-8477-4adf-b430-f6a558ccc8ff>', 'dump': 'CC-MAIN-2013-20', 'url': 'http://%[email protected]/Content/Press-Release/2012/How-AP-reported-in-all-formats-from-tornado-stricken-regions', 'date': '2013-05-18T05:48:54Z', 'file_path': 's3://commoncrawl/crawl-data/CC-MAIN-2013-20/segments/1368696381249/warc/CC-MAIN-20130516092621-00000-ip-10-60-113-184.ec2.internal.warc.gz', 'language': 'en', 'language_score': 0.9721424579620361, 'token_count': 717}\nterminate called without an active exception\nAborted (core dumped)\nmilovidov@milovidov-pc:~/work/datasets$ \npython3 --version\nPython 3.10.12\n```",
"Thank you @alexey-milovidov for the details, was able to reproduce the issue.\n\nFollowing is a preliminary analysis which would help to further isolate the issue:\nOn local: \n- For alternate datasets e.g. `speed/english_quotes_paraphrase` instead of `HuggingFaceFW/fineweb` the code works\n- Multiple calls of `print(next(iter(dataset)))` can be performed successfully before the `terminate` is raised, indicating possibility of issue when connection is closed\n\nOn colab:\n- The above code works properly"
] |
3,051,731,207 |
add check if repo exists for dataset uploading
|
open
|
Currently, I'm reuploading datasets for [`MTEB`](https://github.com/embeddings-benchmark/mteb/). Some of them have many splits (more than 20), and I'm encountering the error:
`Too many requests for https://huggingface.co/datasets/repo/create`.
It seems that this issue occurs because the dataset tries to recreate itself every time a split is uploaded. To resolve this, I've added a check to ensure that if the dataset already exists, it won't attempt to recreate it.
| 2025-05-09T10:27:00 | 2025-06-09T14:39:23 | null |
https://github.com/huggingface/datasets/pull/7565
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7565",
"html_url": "https://github.com/huggingface/datasets/pull/7565",
"diff_url": "https://github.com/huggingface/datasets/pull/7565.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7565.patch",
"merged_at": null
}
| 7,565 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7565). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.",
"@lhoestq Can you review, please? I don't think that errors in CI are related to my changes"
] |
3,049,275,226 |
Implementation of iteration over values of a column in an IterableDataset object
|
closed
|
Refers to [this issue](https://github.com/huggingface/datasets/issues/7381).
| 2025-05-08T14:59:22 | 2025-05-19T12:15:02 | 2025-05-19T12:15:02 |
https://github.com/huggingface/datasets/pull/7564
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7564",
"html_url": "https://github.com/huggingface/datasets/pull/7564",
"diff_url": "https://github.com/huggingface/datasets/pull/7564.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7564.patch",
"merged_at": "2025-05-19T12:15:02"
}
| 7,564 | true |
[
"A couple of questions:\r\n1. I've noticed two strange things: 1) \"Around 80% of the final dataset is made of the `en_dataset`\" in https://huggingface.co/docs/datasets/stream, 2) \"Click on \"Pull request\" to send your to the project maintainers\" in https://github.com/huggingface/datasets/blob/main/CONTRIBUTING.md Are `en_dataset` and \"your [???]\" typos? If so, I can fix them in this PR.\r\n2. Should I update https://huggingface.co/docs/datasets/stream or https://huggingface.co/docs/datasets/access#iterabledataset to include the new feature?",
"Great ! and chained indexing was easy indeed, thanks :)\r\n\r\nregarding your questions:\r\n\r\n> I've noticed two strange things: 1) \"Around 80% of the final dataset is made of the en_dataset\" in https://huggingface.co/docs/datasets/stream, 2) \"Click on \"Pull request\" to send your to the project maintainers\" in https://github.com/huggingface/datasets/blob/main/CONTRIBUTING.md Are en_dataset and \"your [???]\" typos? If so, I can fix them in this PR.\r\n\r\nOh good catch, both should be fixed indeed. Feel free to open a new PR for those docs fixes\r\n\r\n> Should I update https://huggingface.co/docs/datasets/stream or https://huggingface.co/docs/datasets/access#iterabledataset to include the new feature?\r\n\r\nYep good idea, I think in both places, since /stream is supposed to be exhaustive, and /access already mentions accessing a specific column for `Dataset`",
"@lhoestq, thank you for the answers!\r\n\r\n> Yep good idea, I think in both places, since /stream is supposed to be exhaustive, and /access already mentions accessing a specific column for Dataset\r\n\r\n👍, I'll try to add something.\r\n\r\nBy the way, do you have any ideas about why the CI pipelines have failed? Essentially, I've already encountered these problems [here](https://github.com/huggingface/datasets/issues/7381#issuecomment-2863421974).\r\nI think `check_code_quality` has failed due to the usage of `pre-commit`. The problem seems to be the old version of the ruff hook. I've tried `v0.11.8` (the one that was installed with `pip install -e \".[quality]\"`) and `pre-commit` seems to work like `make style` now. However, I don't have any ideas about `pyav` since I don't know what it is...",
"I've updated /stream and /access, please check the style and clarity. By the way, I would like to add `IterableDataset.skip` near `IterableDataset.take` to mimic [slicing](https://huggingface.co/docs/datasets/access/#slicing). What do you think?",
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7564). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,046,351,253 |
set dev version
|
closed
| null | 2025-05-07T15:18:29 | 2025-05-07T15:21:05 | 2025-05-07T15:18:36 |
https://github.com/huggingface/datasets/pull/7563
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7563",
"html_url": "https://github.com/huggingface/datasets/pull/7563",
"diff_url": "https://github.com/huggingface/datasets/pull/7563.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7563.patch",
"merged_at": "2025-05-07T15:18:36"
}
| 7,563 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7563). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,046,339,430 |
release: 3.6.0
|
closed
| null | 2025-05-07T15:15:13 | 2025-05-07T15:17:46 | 2025-05-07T15:15:21 |
https://github.com/huggingface/datasets/pull/7562
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7562",
"html_url": "https://github.com/huggingface/datasets/pull/7562",
"diff_url": "https://github.com/huggingface/datasets/pull/7562.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7562.patch",
"merged_at": "2025-05-07T15:15:20"
}
| 7,562 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7562). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,046,302,653 |
NotImplementedError: <class 'datasets.iterable_dataset.RepeatExamplesIterable'> doesn't implement num_shards yet
|
closed
|
### Describe the bug
When using `.repeat()` on an `IterableDataset`, this error gets thrown. There is [this thread](https://discuss.huggingface.co/t/making-an-infinite-iterabledataset/146192/5) that seems to imply the fix is trivial, but I don't know anything about this codebase, so I'm opening this issue rather than attempting to open a PR.
### Steps to reproduce the bug
1. Create an `IterableDataset`.
2. Call `.repeat(None)` on it.
3. Wrap it in a pytorch `DataLoader`
4. Iterate over it.
### Expected behavior
This should work normally.
### Environment info
datasets: 3.5.0
| 2025-05-07T15:05:42 | 2025-06-05T12:41:30 | 2025-06-05T12:41:30 |
https://github.com/huggingface/datasets/issues/7561
| null | 7,561 | false |
[] |
3,046,265,500 |
fix decoding tests
|
closed
| null | 2025-05-07T14:56:14 | 2025-05-07T14:59:02 | 2025-05-07T14:56:20 |
https://github.com/huggingface/datasets/pull/7560
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7560",
"html_url": "https://github.com/huggingface/datasets/pull/7560",
"diff_url": "https://github.com/huggingface/datasets/pull/7560.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7560.patch",
"merged_at": "2025-05-07T14:56:20"
}
| 7,560 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7560). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,046,177,078 |
fix aiohttp import
|
closed
| null | 2025-05-07T14:31:32 | 2025-05-07T14:34:34 | 2025-05-07T14:31:38 |
https://github.com/huggingface/datasets/pull/7559
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7559",
"html_url": "https://github.com/huggingface/datasets/pull/7559",
"diff_url": "https://github.com/huggingface/datasets/pull/7559.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7559.patch",
"merged_at": "2025-05-07T14:31:38"
}
| 7,559 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7559). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,046,066,628 |
fix regression
|
closed
|
reported in https://github.com/huggingface/datasets/pull/7557 (I just reorganized the condition)
wanted to apply this change to the original PR but github didn't let me apply it directly - merging this one instead
| 2025-05-07T13:56:03 | 2025-05-07T13:58:52 | 2025-05-07T13:56:18 |
https://github.com/huggingface/datasets/pull/7558
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7558",
"html_url": "https://github.com/huggingface/datasets/pull/7558",
"diff_url": "https://github.com/huggingface/datasets/pull/7558.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7558.patch",
"merged_at": "2025-05-07T13:56:18"
}
| 7,558 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7558). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,045,962,076 |
check for empty _formatting
|
closed
|
Fixes a regression from #7553 breaking shuffling of iterable datasets
<img width="884" alt="Screenshot 2025-05-07 at 9 16 52 AM" src="https://github.com/user-attachments/assets/d2f43c5f-4092-4efe-ac31-a32cbd025fe3" />
| 2025-05-07T13:22:37 | 2025-05-07T13:57:12 | 2025-05-07T13:57:12 |
https://github.com/huggingface/datasets/pull/7557
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7557",
"html_url": "https://github.com/huggingface/datasets/pull/7557",
"diff_url": "https://github.com/huggingface/datasets/pull/7557.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7557.patch",
"merged_at": null
}
| 7,557 | true |
[
"Thanks for reporting and for the fix ! I tried to reorganize the condition in your PR but didn't get the right permission so. I ended up merging https://github.com/huggingface/datasets/pull/7558 directly so I can make a release today - I hope you don't mind"
] |
3,043,615,210 |
Add `--merge-pull-request` option for `convert_to_parquet`
|
closed
|
Closes #7527
Note that this implementation **will only merge the last PR in the case that they get split up by `push_to_hub`**. See https://github.com/huggingface/datasets/discussions/7555 for more details.
| 2025-05-06T18:05:05 | 2025-07-18T19:09:10 | 2025-07-18T19:09:10 |
https://github.com/huggingface/datasets/pull/7556
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7556",
"html_url": "https://github.com/huggingface/datasets/pull/7556",
"diff_url": "https://github.com/huggingface/datasets/pull/7556.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7556.patch",
"merged_at": null
}
| 7,556 | true |
[
"This is ready for a review, happy to make any changes. The main question for maintainers is how this should interact with #7555. If my suggestion there is accepted, this PR can be kept as is. If not, more changes are required to merge all the PR parts.",
"Closing since convert to parquet has been removed... https://github.com/huggingface/datasets/pull/7592#issuecomment-3073053138"
] |
3,043,089,844 |
datasets downloads and generates all splits, even though a single split is requested (for dataset with loading script)
|
closed
|
### Describe the bug
`datasets` downloads and generates all splits, even though a single split is requested. [This](https://huggingface.co/datasets/jordiae/exebench) is the dataset in question. It uses a loading script. I am not 100% sure that this is a bug, because maybe with loading scripts `datasets` must actually process all the splits? But I thought loading scripts were designed to avoid this.
### Steps to reproduce the bug
See [this notebook](https://colab.research.google.com/drive/14kcXp_hgcdj-kIzK0bCG6taE-CLZPVvq?usp=sharing)
Or:
```python
from datasets import load_dataset
dataset = load_dataset('jordiae/exebench', split='test_synth', trust_remote_code=True)
```
### Expected behavior
I expected only the `test_synth` split to be downloaded and processed.
### Environment info
- `datasets` version: 3.5.1
- Platform: Linux-6.1.123+-x86_64-with-glibc2.35
- Python version: 3.11.12
- `huggingface_hub` version: 0.30.2
- PyArrow version: 18.1.0
- Pandas version: 2.2.2
- `fsspec` version: 2025.3.0
| 2025-05-06T14:43:38 | 2025-05-07T14:53:45 | 2025-05-07T14:53:44 |
https://github.com/huggingface/datasets/issues/7554
| null | 7,554 | false |
[
"Hi ! there has been some effort on allowing to download only a subset of splits in https://github.com/huggingface/datasets/pull/6832 but no one has been continuing this work so far. This would be a welcomed contribution though\n\nAlso note that loading script are often unoptimized, and we recommend using datasets in standard formats like Parquet instead.\n\nBtw there is a CLI tool to convert a loading script to parquet:\n\n```\ndatasets-cli convert_to_parquet <dataset-name> --trust_remote_code\n```",
"Closing in favor of #6832 "
] |
3,042,953,907 |
Rebatch arrow iterables before formatted iterable
|
closed
|
close https://github.com/huggingface/datasets/issues/7538 and https://github.com/huggingface/datasets/issues/7475
| 2025-05-06T13:59:58 | 2025-05-07T13:17:41 | 2025-05-06T14:03:42 |
https://github.com/huggingface/datasets/pull/7553
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7553",
"html_url": "https://github.com/huggingface/datasets/pull/7553",
"diff_url": "https://github.com/huggingface/datasets/pull/7553.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7553.patch",
"merged_at": "2025-05-06T14:03:41"
}
| 7,553 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7553). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.",
"@lhoestq Our CI found an issue with this changeset causing a regression with shuffling iterable datasets \r\n<img width=\"884\" alt=\"Screenshot 2025-05-07 at 9 16 52 AM\" src=\"https://github.com/user-attachments/assets/bf7d9c7e-cc14-47da-8da6-d1a345992d7c\" />\r\n"
] |
3,040,258,084 |
Enable xet in push to hub
|
closed
|
follows https://github.com/huggingface/huggingface_hub/pull/3035
related to https://github.com/huggingface/datasets/issues/7526
| 2025-05-05T17:02:09 | 2025-05-06T12:42:51 | 2025-05-06T12:42:48 |
https://github.com/huggingface/datasets/pull/7552
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7552",
"html_url": "https://github.com/huggingface/datasets/pull/7552",
"diff_url": "https://github.com/huggingface/datasets/pull/7552.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7552.patch",
"merged_at": "2025-05-06T12:42:48"
}
| 7,552 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7552). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,038,114,928 |
Issue with offline mode and partial dataset cached
|
open
|
### Describe the bug
Hi,
a issue related to #4760 here when loading a single file from a dataset, unable to access it in offline mode afterwards
### Steps to reproduce the bug
```python
import os
# os.environ["HF_HUB_OFFLINE"] = "1"
os.environ["HF_TOKEN"] = "xxxxxxxxxxxxxx"
import datasets
dataset_name = "uonlp/CulturaX"
data_files = "fr/fr_part_00038.parquet"
ds = datasets.load_dataset(dataset_name, split='train', data_files=data_files)
print(f"Dataset loaded : {ds}")
```
Once the file has been cached, I rerun with the HF_HUB_OFFLINE activated an get this error :
```
ValueError: Couldn't find cache for uonlp/CulturaX for config 'default-1e725f978350254e'
Available configs in the cache: ['default-2935e8cdcc21c613']
```
### Expected behavior
Should be able to access the previously cached files
### Environment info
- `datasets` version: 3.2.0
- Platform: Linux-5.4.0-215-generic-x86_64-with-glibc2.31
- Python version: 3.12.0
- `huggingface_hub` version: 0.27.0
- PyArrow version: 19.0.0
- Pandas version: 2.2.2
- `fsspec` version: 2024.3.1
| 2025-05-04T16:49:37 | 2025-05-13T03:18:43 | null |
https://github.com/huggingface/datasets/issues/7551
| null | 7,551 | false |
[
"It seems the problem comes from builder.py / create_config_id()\n\nOn the first call, when the cache is empty we have\n```\nconfig_kwargs = {'data_files': {'train': ['hf://datasets/uonlp/CulturaX@6a8734bc69fefcbb7735f4f9250f43e4cd7a442e/fr/fr_part_00038.parquet']}}\n```\nleading to config_id beeing 'default-2935e8cdcc21c613'\n\nthen, on the second call, \n```\nconfig_kwargs = {'data_files': 'fr/fr_part_00038.parquet'}\n```\nthus explaining why the hash is not the same, despite having the same parameter when calling load_dataset : data_files=\"fr/fr_part_00038.parquet\"",
"Same behavior with version 3.5.1",
"Same issue when loading `google/IndicGenBench_flores_in` with `dataset==2.21.0` and `dataset==3.6.0` .",
"\n\n\n> It seems the problem comes from builder.py / create_config_id()\n> \n> On the first call, when the cache is empty we have\n> \n> ```\n> config_kwargs = {'data_files': {'train': ['hf://datasets/uonlp/CulturaX@6a8734bc69fefcbb7735f4f9250f43e4cd7a442e/fr/fr_part_00038.parquet']}}\n> ```\n> \n> leading to config_id beeing 'default-2935e8cdcc21c613'\n> \n> then, on the second call,\n> \n> ```\n> config_kwargs = {'data_files': 'fr/fr_part_00038.parquet'}\n> ```\n> \n> thus explaining why the hash is not the same, despite having the same parameter when calling load_dataset : data_files=\"fr/fr_part_00038.parquet\"\n\n\nI have identified that the issue indeed lies in the `data_files` within `config_kwargs`. \nThe format and prefix of `data_files` differ depending on whether `HF_HUB_OFFLINE` is set, leading to different final `config_id` values. \nWhen I use other datasets without passing the `data_files` parameter, this issue does not occur.\n\nA possible solution might be to standardize the formatting of `data_files` within the `create_config_id` function."
] |
3,037,017,367 |
disable aiohttp depend for python 3.13t free-threading compat
|
closed
| null | 2025-05-03T00:28:18 | 2025-05-03T00:28:24 | 2025-05-03T00:28:24 |
https://github.com/huggingface/datasets/pull/7550
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7550",
"html_url": "https://github.com/huggingface/datasets/pull/7550",
"diff_url": "https://github.com/huggingface/datasets/pull/7550.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7550.patch",
"merged_at": null
}
| 7,550 | true |
[] |
3,036,272,015 |
TypeError: Couldn't cast array of type string to null on webdataset format dataset
|
open
|
### Describe the bug
```python
from datasets import load_dataset
dataset = load_dataset("animetimm/danbooru-wdtagger-v4-w640-ws-30k")
```
got
```
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/arrow_writer.py", line 626, in write_batch
arrays.append(pa.array(typed_sequence))
File "pyarrow/array.pxi", line 255, in pyarrow.lib.array
File "pyarrow/array.pxi", line 117, in pyarrow.lib._handle_arrow_array_protocol
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/arrow_writer.py", line 258, in __arrow_array__
out = cast_array_to_feature(
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/table.py", line 1798, in wrapper
return func(array, *args, **kwargs)
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/table.py", line 2006, in cast_array_to_feature
arrays = [
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/table.py", line 2007, in <listcomp>
_c(array.field(name) if name in array_fields else null_array, subfeature)
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/table.py", line 1798, in wrapper
return func(array, *args, **kwargs)
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/table.py", line 2066, in cast_array_to_feature
casted_array_values = _c(array.values, feature.feature)
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/table.py", line 1798, in wrapper
return func(array, *args, **kwargs)
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/table.py", line 2103, in cast_array_to_feature
return array_cast(
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/table.py", line 1798, in wrapper
return func(array, *args, **kwargs)
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/table.py", line 1949, in array_cast
raise TypeError(f"Couldn't cast array of type {_short_str(array.type)} to {_short_str(pa_type)}")
TypeError: Couldn't cast array of type string to null
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/load.py", line 2084, in load_dataset
builder_instance.download_and_prepare(
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/builder.py", line 925, in download_and_prepare
self._download_and_prepare(
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/builder.py", line 1649, in _download_and_prepare
super()._download_and_prepare(
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/builder.py", line 1001, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/builder.py", line 1487, in _prepare_split
for job_id, done, content in self._prepare_split_single(
File "/home/ubuntu/miniconda3/lib/python3.10/site-packages/datasets/builder.py", line 1644, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset
```
`datasets==3.5.1` whats wrong
its inner json structure is like
```yaml
features:
- name: "image"
dtype: "image"
- name: "json.id"
dtype: "string"
- name: "json.width"
dtype: "int32"
- name: "json.height"
dtype: "int32"
- name: "json.rating"
sequence:
dtype: "string"
- name: "json.general_tags"
sequence:
dtype: "string"
- name: "json.character_tags"
sequence:
dtype: "string"
```
i'm 100% sure all the jsons satisfies the abovementioned format.
### Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = load_dataset("animetimm/danbooru-wdtagger-v4-w640-ws-30k")
```
### Expected behavior
load the dataset successfully, with the abovementioned json format and webp images
### Environment info
Copy-and-paste the text below in your GitHub issue.
- `datasets` version: 3.5.1
- Platform: Linux-6.8.0-52-generic-x86_64-with-glibc2.35
- Python version: 3.10.16
- `huggingface_hub` version: 0.30.2
- PyArrow version: 20.0.0
- Pandas version: 2.2.3
- `fsspec` version: 2025.3.0
| 2025-05-02T15:18:07 | 2025-05-02T15:37:05 | null |
https://github.com/huggingface/datasets/issues/7549
| null | 7,549 | false |
[
"seems to get fixed by explicitly adding `dataset_infos.json` like this\n\n```json\n{\n \"default\": {\n \"description\": \"Image dataset with tags and ratings\",\n \"citation\": \"\",\n \"homepage\": \"\",\n \"license\": \"\",\n \"features\": {\n \"image\": {\n \"dtype\": \"image\",\n \"_type\": \"Image\"\n },\n \"json\": {\n \"id\": {\n \"dtype\": \"int32\",\n \"_type\": \"Value\"\n },\n \"width\": {\n \"dtype\": \"int32\",\n \"_type\": \"Value\"\n },\n \"height\": {\n \"dtype\": \"int32\",\n \"_type\": \"Value\"\n },\n \"rating\": {\n \"feature\": {\n \"dtype\": \"string\",\n \"_type\": \"Value\"\n },\n \"_type\": \"Sequence\"\n },\n \"general_tags\": {\n \"feature\": {\n \"dtype\": \"string\",\n \"_type\": \"Value\"\n },\n \"_type\": \"Sequence\"\n },\n \"character_tags\": {\n \"feature\": {\n \"dtype\": \"string\",\n \"_type\": \"Value\"\n },\n \"_type\": \"Sequence\"\n }\n }\n },\n \"builder_name\": \"webdataset\",\n \"config_name\": \"default\",\n \"version\": {\n \"version_str\": \"1.0.0\",\n \"description\": null,\n \"major\": 1,\n \"minor\": 0,\n \"patch\": 0\n }\n }\n}\n\n```\n\nwill close this issue if no further issues found"
] |
3,035,568,851 |
Python 3.13t (free threads) Compat
|
open
|
### Describe the bug
Cannot install `datasets` under `python 3.13t` due to dependency on `aiohttp` and aiohttp cannot be built for free-threading python.
The `free threading` support issue in `aiothttp` is active since August 2024! Ouch.
https://github.com/aio-libs/aiohttp/issues/8796#issue-2475941784
`pip install dataset`
```bash
(vm313t) root@gpu-base:~/GPTQModel# pip install datasets
WARNING: Retrying (Retry(total=4, connect=None, read=None, redirect=None, status=None)) after connection broken by 'ReadTimeoutError("HTTPSConnectionPool(host='pypi.org', port=443): Read timed out. (read timeout=15)")': /simple/datasets/
Collecting datasets
Using cached datasets-3.5.1-py3-none-any.whl.metadata (19 kB)
Requirement already satisfied: filelock in /root/vm313t/lib/python3.13t/site-packages (from datasets) (3.18.0)
Requirement already satisfied: numpy>=1.17 in /root/vm313t/lib/python3.13t/site-packages (from datasets) (2.2.5)
Collecting pyarrow>=15.0.0 (from datasets)
Using cached pyarrow-20.0.0-cp313-cp313t-manylinux_2_28_x86_64.whl.metadata (3.3 kB)
Collecting dill<0.3.9,>=0.3.0 (from datasets)
Using cached dill-0.3.8-py3-none-any.whl.metadata (10 kB)
Collecting pandas (from datasets)
Using cached pandas-2.2.3-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (89 kB)
Requirement already satisfied: requests>=2.32.2 in /root/vm313t/lib/python3.13t/site-packages (from datasets) (2.32.3)
Requirement already satisfied: tqdm>=4.66.3 in /root/vm313t/lib/python3.13t/site-packages (from datasets) (4.67.1)
Collecting xxhash (from datasets)
Using cached xxhash-3.5.0-cp313-cp313t-linux_x86_64.whl
Collecting multiprocess<0.70.17 (from datasets)
Using cached multiprocess-0.70.16-py312-none-any.whl.metadata (7.2 kB)
Collecting fsspec<=2025.3.0,>=2023.1.0 (from fsspec[http]<=2025.3.0,>=2023.1.0->datasets)
Using cached fsspec-2025.3.0-py3-none-any.whl.metadata (11 kB)
Collecting aiohttp (from datasets)
Using cached aiohttp-3.11.18.tar.gz (7.7 MB)
Installing build dependencies ... done
Getting requirements to build wheel ... done
Preparing metadata (pyproject.toml) ... done
Requirement already satisfied: huggingface-hub>=0.24.0 in /root/vm313t/lib/python3.13t/site-packages (from datasets) (0.30.2)
Requirement already satisfied: packaging in /root/vm313t/lib/python3.13t/site-packages (from datasets) (25.0)
Requirement already satisfied: pyyaml>=5.1 in /root/vm313t/lib/python3.13t/site-packages (from datasets) (6.0.2)
Collecting aiohappyeyeballs>=2.3.0 (from aiohttp->datasets)
Using cached aiohappyeyeballs-2.6.1-py3-none-any.whl.metadata (5.9 kB)
Collecting aiosignal>=1.1.2 (from aiohttp->datasets)
Using cached aiosignal-1.3.2-py2.py3-none-any.whl.metadata (3.8 kB)
Collecting attrs>=17.3.0 (from aiohttp->datasets)
Using cached attrs-25.3.0-py3-none-any.whl.metadata (10 kB)
Collecting frozenlist>=1.1.1 (from aiohttp->datasets)
Using cached frozenlist-1.6.0-cp313-cp313t-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (16 kB)
Collecting multidict<7.0,>=4.5 (from aiohttp->datasets)
Using cached multidict-6.4.3-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (5.3 kB)
Collecting propcache>=0.2.0 (from aiohttp->datasets)
Using cached propcache-0.3.1-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (10 kB)
Collecting yarl<2.0,>=1.17.0 (from aiohttp->datasets)
Using cached yarl-1.20.0-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (72 kB)
Requirement already satisfied: idna>=2.0 in /root/vm313t/lib/python3.13t/site-packages (from yarl<2.0,>=1.17.0->aiohttp->datasets) (3.10)
Requirement already satisfied: typing-extensions>=3.7.4.3 in /root/vm313t/lib/python3.13t/site-packages (from huggingface-hub>=0.24.0->datasets) (4.13.2)
Requirement already satisfied: charset-normalizer<4,>=2 in /root/vm313t/lib/python3.13t/site-packages (from requests>=2.32.2->datasets) (3.4.1)
Requirement already satisfied: urllib3<3,>=1.21.1 in /root/vm313t/lib/python3.13t/site-packages (from requests>=2.32.2->datasets) (2.4.0)
Requirement already satisfied: certifi>=2017.4.17 in /root/vm313t/lib/python3.13t/site-packages (from requests>=2.32.2->datasets) (2025.4.26)
Collecting python-dateutil>=2.8.2 (from pandas->datasets)
Using cached python_dateutil-2.9.0.post0-py2.py3-none-any.whl.metadata (8.4 kB)
Collecting pytz>=2020.1 (from pandas->datasets)
Using cached pytz-2025.2-py2.py3-none-any.whl.metadata (22 kB)
Collecting tzdata>=2022.7 (from pandas->datasets)
Using cached tzdata-2025.2-py2.py3-none-any.whl.metadata (1.4 kB)
Collecting six>=1.5 (from python-dateutil>=2.8.2->pandas->datasets)
Using cached six-1.17.0-py2.py3-none-any.whl.metadata (1.7 kB)
Using cached datasets-3.5.1-py3-none-any.whl (491 kB)
Using cached dill-0.3.8-py3-none-any.whl (116 kB)
Using cached fsspec-2025.3.0-py3-none-any.whl (193 kB)
Using cached multiprocess-0.70.16-py312-none-any.whl (146 kB)
Using cached multidict-6.4.3-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (220 kB)
Using cached yarl-1.20.0-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (404 kB)
Using cached aiohappyeyeballs-2.6.1-py3-none-any.whl (15 kB)
Using cached aiosignal-1.3.2-py2.py3-none-any.whl (7.6 kB)
Using cached attrs-25.3.0-py3-none-any.whl (63 kB)
Using cached frozenlist-1.6.0-cp313-cp313t-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl (385 kB)
Using cached propcache-0.3.1-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (282 kB)
Using cached pyarrow-20.0.0-cp313-cp313t-manylinux_2_28_x86_64.whl (42.2 MB)
Using cached pandas-2.2.3-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (11.9 MB)
Using cached python_dateutil-2.9.0.post0-py2.py3-none-any.whl (229 kB)
Using cached pytz-2025.2-py2.py3-none-any.whl (509 kB)
Using cached six-1.17.0-py2.py3-none-any.whl (11 kB)
Using cached tzdata-2025.2-py2.py3-none-any.whl (347 kB)
Building wheels for collected packages: aiohttp
Building wheel for aiohttp (pyproject.toml) ... error
error: subprocess-exited-with-error
× Building wheel for aiohttp (pyproject.toml) did not run successfully.
│ exit code: 1
╰─> [156 lines of output]
*********************
* Accelerated build *
*********************
/tmp/pip-build-env-wjqi8_7w/overlay/lib/python3.13t/site-packages/setuptools/dist.py:759: SetuptoolsDeprecationWarning: License classifiers are deprecated.
!!
********************************************************************************
Please consider removing the following classifiers in favor of a SPDX license expression:
License :: OSI Approved :: Apache Software License
See https://packaging.python.org/en/latest/guides/writing-pyproject-toml/#license for details.
********************************************************************************
!!
self._finalize_license_expression()
running bdist_wheel
running build
running build_py
creating build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/typedefs.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/http_parser.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/client_reqrep.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/client_ws.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web_app.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/http_websocket.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/resolver.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/tracing.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/http_writer.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/http_exceptions.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/log.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/__init__.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web_runner.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/worker.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/connector.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/client_exceptions.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web_middlewares.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/tcp_helpers.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web_response.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web_server.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web_request.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web_urldispatcher.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web_exceptions.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/formdata.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/streams.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/multipart.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web_routedef.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web_ws.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/payload.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/client_proto.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web_log.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/base_protocol.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/payload_streamer.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/http.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web_fileresponse.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/test_utils.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/client.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/cookiejar.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/compression_utils.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/hdrs.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/helpers.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/pytest_plugin.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/web_protocol.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/abc.py -> build/lib.linux-x86_64-cpython-313t/aiohttp
creating build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket
copying aiohttp/_websocket/__init__.py -> build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket
copying aiohttp/_websocket/writer.py -> build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket
copying aiohttp/_websocket/models.py -> build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket
copying aiohttp/_websocket/reader.py -> build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket
copying aiohttp/_websocket/reader_c.py -> build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket
copying aiohttp/_websocket/helpers.py -> build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket
copying aiohttp/_websocket/reader_py.py -> build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket
running egg_info
writing aiohttp.egg-info/PKG-INFO
writing dependency_links to aiohttp.egg-info/dependency_links.txt
writing requirements to aiohttp.egg-info/requires.txt
writing top-level names to aiohttp.egg-info/top_level.txt
reading manifest file 'aiohttp.egg-info/SOURCES.txt'
reading manifest template 'MANIFEST.in'
warning: no files found matching 'aiohttp' anywhere in distribution
warning: no files found matching '*.pyi' anywhere in distribution
warning: no previously-included files matching '*.pyc' found anywhere in distribution
warning: no previously-included files matching '*.pyd' found anywhere in distribution
warning: no previously-included files matching '*.so' found anywhere in distribution
warning: no previously-included files matching '*.lib' found anywhere in distribution
warning: no previously-included files matching '*.dll' found anywhere in distribution
warning: no previously-included files matching '*.a' found anywhere in distribution
warning: no previously-included files matching '*.obj' found anywhere in distribution
warning: no previously-included files found matching 'aiohttp/*.html'
no previously-included directories found matching 'docs/_build'
adding license file 'LICENSE.txt'
writing manifest file 'aiohttp.egg-info/SOURCES.txt'
copying aiohttp/_cparser.pxd -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/_find_header.pxd -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/_headers.pxi -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/_http_parser.pyx -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/_http_writer.pyx -> build/lib.linux-x86_64-cpython-313t/aiohttp
copying aiohttp/py.typed -> build/lib.linux-x86_64-cpython-313t/aiohttp
creating build/lib.linux-x86_64-cpython-313t/aiohttp/.hash
copying aiohttp/.hash/_cparser.pxd.hash -> build/lib.linux-x86_64-cpython-313t/aiohttp/.hash
copying aiohttp/.hash/_find_header.pxd.hash -> build/lib.linux-x86_64-cpython-313t/aiohttp/.hash
copying aiohttp/.hash/_http_parser.pyx.hash -> build/lib.linux-x86_64-cpython-313t/aiohttp/.hash
copying aiohttp/.hash/_http_writer.pyx.hash -> build/lib.linux-x86_64-cpython-313t/aiohttp/.hash
copying aiohttp/.hash/hdrs.py.hash -> build/lib.linux-x86_64-cpython-313t/aiohttp/.hash
copying aiohttp/_websocket/mask.pxd -> build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket
copying aiohttp/_websocket/mask.pyx -> build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket
copying aiohttp/_websocket/reader_c.pxd -> build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket
creating build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket/.hash
copying aiohttp/_websocket/.hash/mask.pxd.hash -> build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket/.hash
copying aiohttp/_websocket/.hash/mask.pyx.hash -> build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket/.hash
copying aiohttp/_websocket/.hash/reader_c.pxd.hash -> build/lib.linux-x86_64-cpython-313t/aiohttp/_websocket/.hash
running build_ext
building 'aiohttp._websocket.mask' extension
creating build/temp.linux-x86_64-cpython-313t/aiohttp/_websocket
x86_64-linux-gnu-gcc -fno-strict-overflow -Wsign-compare -DNDEBUG -g -O2 -Wall -g -fno-omit-frame-pointer -mno-omit-leaf-frame-pointer -fstack-protector-strong -fstack-clash-protection -Wformat -Werror=format-security -fcf-protection -fPIC -I/root/vm313t/include -I/usr/include/python3.13t -c aiohttp/_websocket/mask.c -o build/temp.linux-x86_64-cpython-313t/aiohttp/_websocket/mask.o
aiohttp/_websocket/mask.c:1864:80: error: unknown type name ‘__pyx_vectorcallfunc’; did you mean ‘vectorcallfunc’?
1864 | static CYTHON_INLINE PyObject *__Pyx_PyVectorcall_FastCallDict(PyObject *func, __pyx_vectorcallfunc vc, PyObject *const *args, size_t nargs, PyObject *kw);
| ^~~~~~~~~~~~~~~~~~~~
| vectorcallfunc
aiohttp/_websocket/mask.c: In function ‘__pyx_f_7aiohttp_10_websocket_4mask__websocket_mask_cython’:
aiohttp/_websocket/mask.c:2905:3: warning: ‘Py_OptimizeFlag’ is deprecated [-Wdeprecated-declarations]
2905 | if (unlikely(__pyx_assertions_enabled())) {
| ^~
In file included from /usr/include/python3.13t/Python.h:76,
from aiohttp/_websocket/mask.c:16:
/usr/include/python3.13t/cpython/pydebug.h:13:37: note: declared here
13 | Py_DEPRECATED(3.12) PyAPI_DATA(int) Py_OptimizeFlag;
| ^~~~~~~~~~~~~~~
aiohttp/_websocket/mask.c: At top level:
aiohttp/_websocket/mask.c:4846:69: error: unknown type name ‘__pyx_vectorcallfunc’; did you mean ‘vectorcallfunc’?
4846 | static PyObject *__Pyx_PyVectorcall_FastCallDict_kw(PyObject *func, __pyx_vectorcallfunc vc, PyObject *const *args, size_t nargs, PyObject *kw)
| ^~~~~~~~~~~~~~~~~~~~
| vectorcallfunc
aiohttp/_websocket/mask.c:4891:80: error: unknown type name ‘__pyx_vectorcallfunc’; did you mean ‘vectorcallfunc’?
4891 | static CYTHON_INLINE PyObject *__Pyx_PyVectorcall_FastCallDict(PyObject *func, __pyx_vectorcallfunc vc, PyObject *const *args, size_t nargs, PyObject *kw)
| ^~~~~~~~~~~~~~~~~~~~
| vectorcallfunc
aiohttp/_websocket/mask.c: In function ‘__Pyx_CyFunction_CallAsMethod’:
aiohttp/_websocket/mask.c:5580:6: error: unknown type name ‘__pyx_vectorcallfunc’; did you mean ‘vectorcallfunc’?
5580 | __pyx_vectorcallfunc vc = __Pyx_CyFunction_func_vectorcall(cyfunc);
| ^~~~~~~~~~~~~~~~~~~~
| vectorcallfunc
aiohttp/_websocket/mask.c:1954:45: warning: initialization of ‘int’ from ‘vectorcallfunc’ {aka ‘struct _object * (*)(struct _object *, struct _object * const*, long unsigned int, struct _object *)’} makes integer from pointer without a cast [-Wint-conversion]
1954 | #define __Pyx_CyFunction_func_vectorcall(f) (((PyCFunctionObject*)f)->vectorcall)
| ^
aiohttp/_websocket/mask.c:5580:32: note: in expansion of macro ‘__Pyx_CyFunction_func_vectorcall’
5580 | __pyx_vectorcallfunc vc = __Pyx_CyFunction_func_vectorcall(cyfunc);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
aiohttp/_websocket/mask.c:5583:16: warning: implicit declaration of function ‘__Pyx_PyVectorcall_FastCallDict’ [-Wimplicit-function-declaration]
5583 | return __Pyx_PyVectorcall_FastCallDict(func, vc, &PyTuple_GET_ITEM(args, 0), (size_t)PyTuple_GET_SIZE(args), kw);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
aiohttp/_websocket/mask.c:5583:16: warning: returning ‘int’ from a function with return type ‘PyObject *’ {aka ‘struct _object *’} makes pointer from integer without a cast [-Wint-conversion]
5583 | return __Pyx_PyVectorcall_FastCallDict(func, vc, &PyTuple_GET_ITEM(args, 0), (size_t)PyTuple_GET_SIZE(args), kw);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
error: command '/usr/bin/x86_64-linux-gnu-gcc' failed with exit code 1
[end of output]
note: This error originates from a subprocess, and is likely not a problem with pip.
ERROR: Failed building wheel for aiohttp
Failed to build aiohttp
ERROR: Failed to build installable wheels for some pyproject.toml based projects (aiohttp)
```
### Steps to reproduce the bug
See above
### Expected behavior
Install
### Environment info
Ubuntu 24.04
| 2025-05-02T09:20:09 | 2025-05-12T15:11:32 | null |
https://github.com/huggingface/datasets/issues/7548
| null | 7,548 | false |
[
"Update: `datasets` use `aiohttp` for data streaming and from what I understand data streaming is useful for large datasets that do not fit in memory and/or multi-modal datasets like image/audio where you only what the actual binary bits to fed in as needed. \n\nHowever, there are also many cases where aiohttp will never be used. Text datasets that are not huge, relative to machine spec, and non-multi-modal datasets. \n\nGetting `aiohttp` fixed for `free threading` appeals to be a large task that is not going to be get done in a quick manner. It may be faster to make `aiohttp` optional and not forced build. Otherwise, testing python 3.13t is going to be a painful install. \n\nI have created a fork/branch that temp disables aiohttp import so non-streaming usage of datasets can be tested under python 3.13.t:\n\nhttps://github.com/Qubitium/datasets/tree/disable-aiohttp-depend",
"We are mostly relying on `huggingface_hub` which uses `requests` to stream files from Hugging Face, so maybe we can move aiohttp to optional dependencies now. Would it solve your issue ? Btw what do you think of `datasets` in the free-threading setting ?",
"> We are mostly relying on `huggingface_hub` which uses `requests` to stream files from Hugging Face, so maybe we can move aiohttp to optional dependencies now. Would it solve your issue ? Btw what do you think of `datasets` in the free-threading setting ?\n\nI am testing transformers + dataset (simple text dataset usage) + GPTQModel for quantization and there were no issues encountered with python 3.13t but my test-case is the base-bare minimal test-case since dataset is not sharded, fully in-memory, text-only, small, not used for training. \n\nOn the technical side, dataset is almost always 100% read-only so there should be zero locking issues but I have not checked the dataset internals so there may be cases where streaming, sharding, and/or cases where datset memory/states are updated needs a per dataset `threading.lock`. \n\nSo yes, making `aiohttp` optional will definitely solve my issue. There is also a companion (datasets and tokenizers usually go hand-in-hand) issue with `Tokenizers` as well but that's simple enough with package version update: https://github.com/huggingface/tokenizers/pull/1774\n",
"Ok I see ! Anyway feel free to edit the setup.py to move aiohttp to optional (tests) dependencies and open a PR, we can run the CI to see if it's ok as a change",
"actually there is https://github.com/huggingface/datasets/pull/7294/ already, let's see if we can merge it",
"wouldn't it be the good reason to switch to `httpx`? 😄 (would require slightly more work, short term agree with https://github.com/huggingface/datasets/issues/7548#issuecomment-2854405923)",
"I made `aiohttp` optional in `datasets` 3.6.0 :)\n\n`datasets` doesn't use it directly anyway, it's only used when someone wants to download files from HTTP URLs outside of HF"
] |
3,034,830,291 |
Avoid global umask for setting file mode.
|
closed
|
This PR updates the method for setting the permissions on `cache_path` after calling `shutil.move`. The call to `shutil.move` may not preserve permissions if the source and destination are on different filesystems. Reading and resetting umask can cause race conditions, so directly read what permissions were set for the `temp_file` instead.
This fixes https://github.com/huggingface/datasets/issues/7536.
| 2025-05-01T22:24:24 | 2025-05-06T13:05:00 | 2025-05-06T13:05:00 |
https://github.com/huggingface/datasets/pull/7547
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7547",
"html_url": "https://github.com/huggingface/datasets/pull/7547",
"diff_url": "https://github.com/huggingface/datasets/pull/7547.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7547.patch",
"merged_at": "2025-05-06T13:05:00"
}
| 7,547 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7547). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,034,018,298 |
Large memory use when loading large datasets to a ZFS pool
|
closed
|
### Describe the bug
When I load large parquet based datasets from the hub like `MLCommons/peoples_speech` using `load_dataset`, all my memory (500GB) is used and isn't released after loading, meaning that the process is terminated by the kernel if I try to load an additional dataset. This makes it impossible to train models using multiple large datasets.
### Steps to reproduce the bug
`uv run --with datasets==3.5.1 python`
```python
from datasets import load_dataset
load_dataset('MLCommons/peoples_speech', 'clean')
load_dataset('mozilla-foundation/common_voice_17_0', 'en')
```
### Expected behavior
I would expect that a lot less than 500GB of RAM would be required to load the dataset, or at least that the RAM usage would be cleared as soon as the dataset is loaded (and thus reside as a memory mapped file) such that other datasets can be loaded.
### Environment info
I am currently using the latest datasets==3.5.1 but I have had the same problem with multiple other versions.
| 2025-05-01T14:43:47 | 2025-05-13T13:30:09 | 2025-05-13T13:29:53 |
https://github.com/huggingface/datasets/issues/7546
| null | 7,546 | false |
[
"Hi ! datasets are memory mapped from disk, so they don't fill out your RAM. Not sure what's the source of your memory issue.\n\nWhat kind of system are you using ? and what kind of disk ?",
"Well, the fact of the matter is that my RAM is getting filled out by running the given example, as shown in [this video](https://streamable.com/usb0ql).\n\nMy system is a GPU server running Ubuntu. The disk is a SATA SSD attached to the server using a backplane. It is formatted with ZFS, mounted in /cache, and my HF_HOME is set to /cache/hf\n\nI really need this fixed, so I am more than willing to test out various suggestions you might have, or write a PR if we can figure out what is going on.",
"I'm not super familiar with ZFS, but it looks like it loads the data in memory when the files are memory mapped, which is an issue.\n\nMaybe it's a caching mechanism ? Since `datasets` accesses every memory mapped file to read a small part (the metadata of the arrow record batches), maybe ZFS brings the whole files in memory for quicker subsequent reads. This is an antipattern when it comes to lazy loading datasets of that size though",
"This is the answer.\n\nI tried changing my HF_HOME to an NFS share, and no RAM is then consumed loading the dataset.\n\nI will try to see if I can find a way to configure the ZFS pool to not cache the files (disabling the ARC/primary cache didn't work), and if I do write the solution in this issue. If I can't I guess I have to reformat my cache drive."
] |
3,031,617,547 |
Networked Pull Through Cache
|
open
|
### Feature request
Introduce a HF_DATASET_CACHE_NETWORK_LOCATION configuration (e.g. an environment variable) together with a companion network cache service.
Enable a three-tier cache lookup for datasets:
1. Local on-disk cache
2. Configurable network cache proxy
3. Official Hugging Face Hub
### Motivation
- Distributed training & ephemeral jobs: In high-performance or containerized clusters, relying solely on a local disk cache either becomes a streaming bottleneck or incurs a heavy cold-start penalty as each job must re-download datasets.
- Traffic & cost reduction: A pull-through network cache lets multiple consumers share a common cache layer, reducing duplicate downloads from the Hub and lowering egress costs.
- Better streaming adoption: By offloading repeat dataset pulls to a locally managed cache proxy, streaming workloads can achieve higher throughput and more predictable latency.
- Proven pattern: Similar proxy-cache solutions (e.g. Harbor’s Proxy Cache for Docker images) have demonstrated reliability and performance at scale: https://goharbor.io/docs/2.1.0/administration/configure-proxy-cache/
### Your contribution
I’m happy to draft the initial PR for adding HF_DATASET_CACHE_NETWORK_LOCATION support in datasets and sketch out a minimal cache-service prototype.
I have limited bandwidth so I would be looking for collaborators if anyone else is interested.
| 2025-04-30T15:16:33 | 2025-04-30T15:16:33 | null |
https://github.com/huggingface/datasets/issues/7545
| null | 7,545 | false |
[] |
3,027,024,285 |
Add try_original_type to DatasetDict.map
|
closed
|
This PR resolves #7472 for DatasetDict
The previously merged PR #7483 added `try_original_type` to ArrowDataset, but DatasetDict misses `try_original_type`
Cc: @lhoestq
| 2025-04-29T04:39:44 | 2025-05-05T14:42:49 | 2025-05-05T14:42:49 |
https://github.com/huggingface/datasets/pull/7544
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7544",
"html_url": "https://github.com/huggingface/datasets/pull/7544",
"diff_url": "https://github.com/huggingface/datasets/pull/7544.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7544.patch",
"merged_at": "2025-05-05T14:42:49"
}
| 7,544 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7544). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.",
"Sure! I just committed the changes",
"@lhoestq \r\nLet me know if there are other things to do before merge or other places to add `try_original_type` argument "
] |
3,026,867,706 |
The memory-disk mapping failure issue of the map function（resolved, but there are some suggestions.）
|
closed
|
### Describe the bug
## bug
When the map function processes a large dataset, it temporarily stores the data in a cache file on the disk. After the data is stored, the memory occupied by it is released. Therefore, when using the map function to process a large-scale dataset, only a dataset space of the size of `writer_batch_size` will be occupied in memory.
However, I found that the map function does not actually reduce memory usage when I used it. At first, I thought there was a bug in the program, causing a memory leak—meaning the memory was not released after the data was stored in the cache. But later, I used a Linux command to check for recently modified files during program execution and found that no new files were created or modified. This indicates that the program did not store the dataset in the disk cache.
## bug solved
After modifying the parameters of the map function multiple times, I discovered the `cache_file_name` parameter. By changing it, the cache file can be stored in the specified directory. After making this change, I noticed that the cache file appeared. Initially, I found this quite incredible, but then I wondered if the cache file might have failed to be stored in a certain folder. This could be related to the fact that I don't have root privileges.
So, I delved into the source code of the map function to find out where the cache file would be stored by default. Eventually, I found the function `def _get_cache_file_path(self, fingerprint):`, which automatically generates the storage path for the cache file. The output was as follows: `/tmp/hf_datasets-j5qco9ug/cache-f2830487643b9cc2.arrow`. My hypothesis was confirmed: the lack of root privileges indeed prevented the cache file from being stored, which in turn prevented the release of memory. Therefore, changing the storage location to a folder where I have write access resolved the issue.
### Steps to reproduce the bug
my code
`train_data = train_data.map(process_fun, remove_columns=['image_name', 'question_type', 'concern', 'question', 'candidate_answers', 'answer'])`
### Expected behavior
Although my bug has been resolved, it still took me nearly a week to search for relevant information and debug the program. However, if a warning or error message about insufficient cache file write permissions could be provided during program execution, I might have been able to identify the cause more quickly. Therefore, I hope this aspect can be improved. I am documenting this bug here so that friends who encounter similar issues can solve their problems in a timely manner.
### Environment info
python: 3.10.15
datasets: 3.5.0
| 2025-04-29T03:04:59 | 2025-04-30T02:22:17 | 2025-04-30T02:22:17 |
https://github.com/huggingface/datasets/issues/7543
| null | 7,543 | false |
[] |
3,025,054,630 |
set dev version
|
closed
| null | 2025-04-28T14:03:48 | 2025-04-28T14:08:37 | 2025-04-28T14:04:00 |
https://github.com/huggingface/datasets/pull/7542
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7542",
"html_url": "https://github.com/huggingface/datasets/pull/7542",
"diff_url": "https://github.com/huggingface/datasets/pull/7542.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7542.patch",
"merged_at": "2025-04-28T14:04:00"
}
| 7,542 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7542). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,025,045,919 |
release: 3.5.1
|
closed
| null | 2025-04-28T14:00:59 | 2025-04-28T14:03:38 | 2025-04-28T14:01:54 |
https://github.com/huggingface/datasets/pull/7541
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7541",
"html_url": "https://github.com/huggingface/datasets/pull/7541",
"diff_url": "https://github.com/huggingface/datasets/pull/7541.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7541.patch",
"merged_at": "2025-04-28T14:01:54"
}
| 7,541 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7541). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,024,862,966 |
support pyarrow 20
|
closed
|
fix
```
TypeError: ArrayExtensionArray.to_pylist() got an unexpected keyword argument 'maps_as_pydicts'
```
| 2025-04-28T13:01:11 | 2025-04-28T13:23:53 | 2025-04-28T13:23:52 |
https://github.com/huggingface/datasets/pull/7540
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7540",
"html_url": "https://github.com/huggingface/datasets/pull/7540",
"diff_url": "https://github.com/huggingface/datasets/pull/7540.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7540.patch",
"merged_at": "2025-04-28T13:23:52"
}
| 7,540 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7540). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,023,311,163 |
Fix IterableDataset state_dict shard_example_idx counting
|
closed
|
# Fix IterableDataset's state_dict shard_example_idx reporting
## Description
This PR fixes issue #7475 where the `shard_example_idx` value in `IterableDataset`'s `state_dict()` always equals the number of samples in a shard, even if only a few examples have been consumed.
The issue is in the `_iter_arrow` method of the `ArrowExamplesIterable` class where it updates the `shard_example_idx` state by the full length of the batch (`len(pa_table)`) even when we're only partway through processing the examples.
## Changes
Modified the `_iter_arrow` method of `ArrowExamplesIterable` to:
1. Track the actual number of examples processed
2. Only increment the `shard_example_idx` by the number of examples actually yielded
3. Handle partial batches correctly
## How to Test
I've included a simple test case that demonstrates the fix:
```python
from datasets import Dataset
# Create a test dataset
ds = Dataset.from_dict({"a": range(6)}).to_iterable_dataset(num_shards=1)
# Iterate through part of the dataset
for idx, example in enumerate(ds):
print(example)
if idx == 2: # Stop after 3 examples (0, 1, 2)
state_dict = ds.state_dict()
print("Checkpoint state_dict:", state_dict)
break
# Before the fix, the output would show shard_example_idx: 6
# After the fix, it shows shard_example_idx: 3, correctly reflecting the 3 processed examples
```
## Implementation Details
1. Added logic to track the number of examples actually seen in the current shard
2. Modified the state update to only count examples actually yielded
3. Improved handling of partial batches and skipped examples
This fix ensures that checkpointing and resuming works correctly with exactly the expected number of examples, rather than skipping ahead to the end of the batch.
| 2025-04-27T20:41:18 | 2025-05-06T14:24:25 | 2025-05-06T14:24:24 |
https://github.com/huggingface/datasets/pull/7539
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7539",
"html_url": "https://github.com/huggingface/datasets/pull/7539",
"diff_url": "https://github.com/huggingface/datasets/pull/7539.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7539.patch",
"merged_at": null
}
| 7,539 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7539). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.",
"Hi ! FYI I made a PR to fix https://github.com/huggingface/datasets/issues/7538 and it also fixed https://github.com/huggingface/datasets/issues/7475, so if I'm not mistaken this PR is not needed anymore"
] |
3,023,280,056 |
`IterableDataset` drops samples when resuming from a checkpoint
|
closed
|
When resuming from a checkpoint, `IterableDataset` will drop samples if `num_shards % world_size == 0` and the underlying example supports `iter_arrow` and needs to be formatted.
In that case, the `FormattedExamplesIterable` fetches a batch of samples from the child iterable's `iter_arrow` and yields them one by one (after formatting). However, the child increments the `shard_example_idx` counter (in its `iter_arrow`) before returning the batch for the whole batch size, which leads to a portion of samples being skipped if the iteration (of the parent iterable) is stopped mid-batch.
Perhaps one way to avoid this would be by signalling the child iterable which samples (within the chunk) are processed by the parent and which are not, so that it can adjust the `shard_example_idx` counter accordingly. This would also mean the chunk needs to be sliced when resuming, but this is straightforward to implement.
The following is a minimal reproducer of the bug:
```python
from datasets import Dataset
from datasets.distributed import split_dataset_by_node
ds = Dataset.from_dict({"n": list(range(24))})
ds = ds.to_iterable_dataset(num_shards=4)
world_size = 4
rank = 0
ds_rank = split_dataset_by_node(ds, rank, world_size)
it = iter(ds_rank)
examples = []
for idx, example in enumerate(it):
examples.append(example)
if idx == 2:
state_dict = ds_rank.state_dict()
break
ds_rank.load_state_dict(state_dict)
it_resumed = iter(ds_rank)
examples_resumed = examples[:]
for example in it:
examples.append(example)
for example in it_resumed:
examples_resumed.append(example)
print("ORIGINAL ITER EXAMPLES:", examples)
print("RESUMED ITER EXAMPLES:", examples_resumed)
```
| 2025-04-27T19:34:49 | 2025-05-06T14:04:05 | 2025-05-06T14:03:42 |
https://github.com/huggingface/datasets/issues/7538
| null | 7,538 | false |
[
"Thanks for reporting ! I fixed the issue using RebatchedArrowExamplesIterable before the formatted iterable"
] |
3,018,792,966 |
`datasets.map(..., num_proc=4)` multi-processing fails
|
open
|
The following code fails in python 3.11+
```python
tokenized_datasets = datasets.map(tokenize_function, batched=True, num_proc=4, remove_columns=["text"])
```
Error log:
```bash
Traceback (most recent call last):
File "/usr/local/lib/python3.12/dist-packages/multiprocess/process.py", line 315, in _bootstrap
self.run()
File "/usr/local/lib/python3.12/dist-packages/multiprocess/process.py", line 108, in run
self._target(*self._args, **self._kwargs)
File "/usr/local/lib/python3.12/dist-packages/multiprocess/pool.py", line 114, in worker
task = get()
^^^^^
File "/usr/local/lib/python3.12/dist-packages/multiprocess/queues.py", line 371, in get
return _ForkingPickler.loads(res)
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/dill/_dill.py", line 327, in loads
return load(file, ignore, **kwds)
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/dill/_dill.py", line 313, in load
return Unpickler(file, ignore=ignore, **kwds).load()
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/dill/_dill.py", line 525, in load
obj = StockUnpickler.load(self)
^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/dill/_dill.py", line 659, in _create_code
if len(args) == 16: return CodeType(*args)
^^^^^^^^^^^^^^^
TypeError: code() argument 13 must be str, not int
```
After upgrading dill to the latest 0.4.0 with "pip install --upgrade dill", it can pass. So it seems that there is a compatibility issue between dill 0.3.4 and python 3.11+, because python 3.10 works fine.
Is the dill deterministic issue mentioned in https://github.com/huggingface/datasets/blob/main/setup.py#L117) still valid? Any plan to unpin?
| 2025-04-25T01:53:47 | 2025-05-06T13:12:08 | null |
https://github.com/huggingface/datasets/issues/7537
| null | 7,537 | false |
[
"related: https://github.com/huggingface/datasets/issues/7510\n\nwe need to do more tests to see if latest `dill` is deterministic"
] |
3,018,425,549 |
[Errno 13] Permission denied: on `.incomplete` file
|
closed
|
### Describe the bug
When downloading a dataset, we frequently hit the below Permission Denied error. This looks to happen (at least) across datasets in HF, S3, and GCS.
It looks like the `temp_file` being passed [here](https://github.com/huggingface/datasets/blob/main/src/datasets/utils/file_utils.py#L412) can sometimes be created with `000` permissions leading to the permission denied error (the user running the code is still the owner of the file). Deleting that particular file and re-running the code with 0 changes will usually succeed.
Is there some race condition happening with the [umask](https://github.com/huggingface/datasets/blob/main/src/datasets/utils/file_utils.py#L416), which is process global, and the [file creation](https://github.com/huggingface/datasets/blob/main/src/datasets/utils/file_utils.py#L404)?
```
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
.venv/lib/python3.12/site-packages/datasets/load.py:2084: in load_dataset
builder_instance.download_and_prepare(
.venv/lib/python3.12/site-packages/datasets/builder.py:925: in download_and_prepare
self._download_and_prepare(
.venv/lib/python3.12/site-packages/datasets/builder.py:1649: in _download_and_prepare
super()._download_and_prepare(
.venv/lib/python3.12/site-packages/datasets/builder.py:979: in _download_and_prepare
split_generators = self._split_generators(dl_manager, **split_generators_kwargs)
.venv/lib/python3.12/site-packages/datasets/packaged_modules/folder_based_builder/folder_based_builder.py:120: in _split_generators
downloaded_files = dl_manager.download(files)
.venv/lib/python3.12/site-packages/datasets/download/download_manager.py:159: in download
downloaded_path_or_paths = map_nested(
.venv/lib/python3.12/site-packages/datasets/utils/py_utils.py:514: in map_nested
_single_map_nested((function, obj, batched, batch_size, types, None, True, None))
.venv/lib/python3.12/site-packages/datasets/utils/py_utils.py:382: in _single_map_nested
return [mapped_item for batch in iter_batched(data_struct, batch_size) for mapped_item in function(batch)]
.venv/lib/python3.12/site-packages/datasets/download/download_manager.py:206: in _download_batched
return thread_map(
.venv/lib/python3.12/site-packages/tqdm/contrib/concurrent.py:69: in thread_map
return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)
.venv/lib/python3.12/site-packages/tqdm/contrib/concurrent.py:51: in _executor_map
return list(tqdm_class(ex.map(fn, *iterables, chunksize=chunksize), **kwargs))
.venv/lib/python3.12/site-packages/tqdm/std.py:1181: in __iter__
for obj in iterable:
../../../_tool/Python/3.12.10/x64/lib/python3.12/concurrent/futures/_base.py:619: in result_iterator
yield _result_or_cancel(fs.pop())
../../../_tool/Python/3.12.10/x64/lib/python3.12/concurrent/futures/_base.py:317: in _result_or_cancel
return fut.result(timeout)
../../../_tool/Python/3.12.10/x64/lib/python3.12/concurrent/futures/_base.py:449: in result
return self.__get_result()
../../../_tool/Python/3.12.10/x64/lib/python3.12/concurrent/futures/_base.py:401: in __get_result
raise self._exception
../../../_tool/Python/3.12.10/x64/lib/python3.12/concurrent/futures/thread.py:59: in run
result = self.fn(*self.args, **self.kwargs)
.venv/lib/python3.12/site-packages/datasets/download/download_manager.py:229: in _download_single
out = cached_path(url_or_filename, download_config=download_config)
.venv/lib/python3.12/site-packages/datasets/utils/file_utils.py:206: in cached_path
output_path = get_from_cache(
.venv/lib/python3.12/site-packages/datasets/utils/file_utils.py:412: in get_from_cache
fsspec_get(url, temp_file, storage_options=storage_options, desc=download_desc, disable_tqdm=disable_tqdm)
.venv/lib/python3.12/site-packages/datasets/utils/file_utils.py:331: in fsspec_get
fs.get_file(path, temp_file.name, callback=callback)
.venv/lib/python3.12/site-packages/fsspec/asyn.py:118: in wrapper
return sync(self.loop, func, *args, **kwargs)
.venv/lib/python3.12/site-packages/fsspec/asyn.py:103: in sync
raise return_result
.venv/lib/python3.12/site-packages/fsspec/asyn.py:56: in _runner
result[0] = await coro
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
self = <s3fs.core.S3FileSystem object at 0x7f27c18b2e70>
rpath = '<my-bucket>/<my-prefix>/img_1.jpg'
lpath = '/home/runner/_work/_temp/hf_cache/downloads/6c97983efa4e24e534557724655df8247a0bd04326cdfc4a95b638c11e78222d.incomplete'
callback = <datasets.utils.file_utils.TqdmCallback object at 0x7f27c00cdbe0>
version_id = None, kwargs = {}
_open_file = <function S3FileSystem._get_file.<locals>._open_file at 0x7f27628d1120>
body = <StreamingBody at 0x7f276344fa80 for ClientResponse at 0x7f27c015fce0>
content_length = 521923, failed_reads = 0, bytes_read = 0
async def _get_file(
self, rpath, lpath, callback=_DEFAULT_CALLBACK, version_id=None, **kwargs
):
if os.path.isdir(lpath):
return
bucket, key, vers = self.split_path(rpath)
async def _open_file(range: int):
kw = self.req_kw.copy()
if range:
kw["Range"] = f"bytes={range}-"
resp = await self._call_s3(
"get_object",
Bucket=bucket,
Key=key,
**version_id_kw(version_id or vers),
**kw,
)
return resp["Body"], resp.get("ContentLength", None)
body, content_length = await _open_file(range=0)
callback.set_size(content_length)
failed_reads = 0
bytes_read = 0
try:
> with open(lpath, "wb") as f0:
E PermissionError: [Errno 13] Permission denied: '/home/runner/_work/_temp/hf_cache/downloads/6c97983efa4e24e534557724655df8247a0bd04326cdfc4a95b638c11e78222d.incomplete'
.venv/lib/python3.12/site-packages/s3fs/core.py:1355: PermissionError
```
### Steps to reproduce the bug
I believe this is a race condition and cannot reliably re-produce it, but it happens fairly frequently in our GitHub Actions tests and can also be re-produced (with lesser frequency) on cloud VMs.
### Expected behavior
The dataset loads properly with no permission denied error.
### Environment info
- `datasets` version: 3.5.0
- Platform: Linux-5.10.0-34-cloud-amd64-x86_64-with-glibc2.31
- Python version: 3.12.10
- `huggingface_hub` version: 0.30.2
- PyArrow version: 19.0.1
- Pandas version: 2.2.3
- `fsspec` version: 2024.12.0
| 2025-04-24T20:52:45 | 2025-05-06T13:05:01 | 2025-05-06T13:05:01 |
https://github.com/huggingface/datasets/issues/7536
| null | 7,536 | false |
[
"It must be an issue with umask being used by multiple threads indeed. Maybe we can try to make a thread safe function to apply the umask (using filelock for example)",
"> It must be an issue with umask being used by multiple threads indeed. Maybe we can try to make a thread safe function to apply the umask (using filelock for example)\n\n@lhoestq is this something which can go in a 3.5.1 release?",
"Yes for sure",
"@lhoestq - can you take a look at https://github.com/huggingface/datasets/pull/7547/?"
] |
3,018,289,872 |
Change dill version in requirements
|
open
|
Change dill version to >=0.3.9,<0.4.5 and check for errors
| 2025-04-24T19:44:28 | 2025-05-19T14:51:29 | null |
https://github.com/huggingface/datasets/pull/7535
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7535",
"html_url": "https://github.com/huggingface/datasets/pull/7535",
"diff_url": "https://github.com/huggingface/datasets/pull/7535.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7535.patch",
"merged_at": null
}
| 7,535 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7535). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
3,017,259,407 |
TensorFlow RaggedTensor Support (batch-level)
|
open
|
### Feature request
Hi,
Currently datasets does not support RaggedTensor output on batch-level.
When building a Object Detection Dataset (with TensorFlow) I need to enable RaggedTensors as that's how BBoxes & classes are expected from the Keras Model POV.
Currently there's a error thrown saying that "Nested Data is not supported".
It'd be very helpful if this was fixed! :)
### Motivation
Enabling Object Detection pipelines for TensorFlow.
### Your contribution
With guidance I'd happily help making the PR.
The current implementation with DataCollator and later enforcing `np.array` is the problematic part (at the end of `np_get_batch` in `tf_utils.py`). As `numpy` don't support "Raggednes"
| 2025-04-24T13:14:52 | 2025-06-30T17:03:39 | null |
https://github.com/huggingface/datasets/issues/7534
| null | 7,534 | false |
[
"Keras doesn't support other inputs other than tf.data.Dataset objects ? it's a bit painful to have to support and maintain this kind of integration\n\nIs there a way to use a `datasets.Dataset` with outputs formatted as tensors / ragged tensors instead ? like in https://huggingface.co/docs/datasets/use_with_tensorflow#dataset-format",
"I'll give it a try when I get the time. But quite sure I already tested the `with_format` approach.\n\nKeras when using TF as backend converts the datasets into `tf.data.Dataset`, much like you do.",
"Hi @Lundez! Thanks for raising this — very valid point, especially for Object Detection use-cases.\n\nYou're right that np_get_batch currently enforces numpy batching, which breaks RaggedTensor support due to its inability to handle nested structures. This likely needs a redesign to allow TensorFlow-native batching in specific formats.\n\nBefore diving into a code change though, could you confirm:\n\nDoes `.with_format(\"tensorflow\")` (without batching) return a `tf.data.Dataset` that works if batching is deferred to `model.fit()`?\n\nHave you tried something like:\n\n```python\ntf_dataset = dataset.with_format(\"tensorflow\").to_tf_dataset(\n columns=[\"image\", \"labels\"],\n label_cols=None,\n batch_size=None # No batching here\n)\nmodel.fit(tf_dataset.batch(BATCH_SIZE)) # Use RaggedTensor batching here\n```\n\nIf this works, it might be worth updating the documentation rather than changing batching logic inside datasets itself.\n\nThat said, happy to explore changes if batching needs to be supported natively for RaggedTensor. Just flagging that it’d require some careful design due to existing numpy assumptions.",
"Hi, we've had to move on for now. \n\nWe have actually also moved to dense tensors to make it possible to xla complie the training. \n\nBut I'll check when I'm back from vacation which is far into the future. \n\nThanks"
] |
3,015,075,086 |
Add custom fingerprint support to `from_generator`
|
open
|
This PR adds `dataset_id_suffix` parameter to 'Dataset.from_generator' function.
`Dataset.from_generator` function passes all of its arguments to `BuilderConfig.create_config_id`, including generator function itself. `BuilderConfig.create_config_id` function tries to hash all the args, which can take a large amount of time or even cause MemoryError if the dataset processed in a generator function is large enough.
This PR allows user to pass a custom fingerprint (`dataset_id_suffix`) to be used as a suffix in a dataset name instead of the one generated by hashing the args.
This PR is a possible solution of #7513
| 2025-04-23T19:31:35 | 2025-07-10T09:29:35 | null |
https://github.com/huggingface/datasets/pull/7533
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7533",
"html_url": "https://github.com/huggingface/datasets/pull/7533",
"diff_url": "https://github.com/huggingface/datasets/pull/7533.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7533.patch",
"merged_at": null
}
| 7,533 | true |
[
"This is great !\r\n\r\nWhat do you think of passing `config_id=` directly to the builder instead of just the suffix ? This would be a power user argument though, or for internal use. And in from_generator the new argument can be `fingerprint=` as in `Dataset.__init__()`\r\n\r\nThe `config_id` can be defined using something like `config_id = \"default-fingerprint=\" + fingerprint`\r\n\r\nI feel ike this could make the Dataset API more coherent if we avoid introducing a new argument while we can juste use `fingerprint=`",
"I looked into this issue and the original cause makes total sense — hashing a large generator is clearly inefficient and fragile for big datasets.\r\n\r\nPR #7533 looks like a robust and flexible solution! It cleanly separates the fingerprinting responsibility by letting users pass `fingerprint=` (now `config_id=`), which avoids hashing heavy objects like generators but still preserves caching logic.\r\n",
"@lhoestq could you please re-review the changes I made?"
] |
3,009,546,204 |
Document the HF_DATASETS_CACHE environment variable in the datasets cache documentation
|
closed
|
This pull request updates the Datasets documentation to include the `HF_DATASETS_CACHE` environment variable. While the current documentation only mentions `HF_HOME` for overriding the default cache directory, `HF_DATASETS_CACHE` is also a supported and useful option for specifying a custom cache location for datasets stored in Arrow format.
This addition is based on the discussion in (https://github.com/huggingface/datasets/issues/7457), where users noted the absence of this variable in the documentation despite its functionality. The update adds a new section to `cache.mdx` that explains how to use `HF_DATASETS_CACHE` with an example.
This change aims to improve clarity and help users better manage their cache directories when working in shared environments or with limited local storage.
Closes #7457.
| 2025-04-22T00:23:13 | 2025-05-06T15:54:38 | 2025-05-06T15:54:38 |
https://github.com/huggingface/datasets/pull/7532
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7532",
"html_url": "https://github.com/huggingface/datasets/pull/7532",
"diff_url": "https://github.com/huggingface/datasets/pull/7532.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7532.patch",
"merged_at": "2025-05-06T15:54:38"
}
| 7,532 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7532). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.",
"Your clarification in your comment at https://github.com/huggingface/datasets/issues/7480#issuecomment-2833640084 sounds great, would you like to update this PR to include it ?",
"Hi @lhoestq, I’ve updated the documentation to reflect the clarifications discussed in #7480. Let me know if anything else is needed!\r\n"
] |
3,008,914,887 |
Deepspeed reward training hangs at end of training with Dataset.from_list
|
open
|
There seems to be a weird interaction between Deepspeed, the Dataset.from_list method and trl's RewardTrainer. On a multi-GPU setup (10 A100s), training always hangs at the very end of training until it times out. The training itself works fine until the end of training and running the same script with Deepspeed on a single GPU works without hangig. The issue persisted across a wide range of Deepspeed configs and training arguments. The issue went away when storing the exact same dataset as a JSON and using `dataset = load_dataset("json", ...)`. Here is my training script:
```python
import pickle
import os
import random
import warnings
import torch
from datasets import load_dataset, Dataset
from transformers import AutoModelForSequenceClassification, AutoTokenizer
from trl import RewardConfig, RewardTrainer, ModelConfig
####################################### Reward model #################################################
# Explicitly set arguments
model_name_or_path = "Qwen/Qwen2.5-1.5B"
output_dir = "Qwen2-0.5B-Reward-LoRA"
per_device_train_batch_size = 2
num_train_epochs = 5
gradient_checkpointing = True
learning_rate = 1.0e-4
logging_steps = 25
eval_strategy = "steps"
eval_steps = 50
max_length = 2048
torch_dtype = "auto"
trust_remote_code = False
model_args = ModelConfig(
model_name_or_path=model_name_or_path,
model_revision=None,
trust_remote_code=trust_remote_code,
torch_dtype=torch_dtype,
lora_task_type="SEQ_CLS", # Make sure task type is seq_cls
)
training_args = RewardConfig(
output_dir=output_dir,
per_device_train_batch_size=per_device_train_batch_size,
num_train_epochs=num_train_epochs,
gradient_checkpointing=gradient_checkpointing,
learning_rate=learning_rate,
logging_steps=logging_steps,
eval_strategy=eval_strategy,
eval_steps=eval_steps,
max_length=max_length,
gradient_checkpointing_kwargs=dict(use_reentrant=False),
center_rewards_coefficient = 0.01,
fp16=False,
bf16=True,
save_strategy="no",
dataloader_num_workers=0,
# deepspeed="./configs/deepspeed_config.json",
)
################
# Model & Tokenizer
################
model_kwargs = dict(
revision=model_args.model_revision,
use_cache=False if training_args.gradient_checkpointing else True,
torch_dtype=model_args.torch_dtype,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path, use_fast=True
)
model = AutoModelForSequenceClassification.from_pretrained(
model_args.model_name_or_path, num_labels=1, trust_remote_code=model_args.trust_remote_code, **model_kwargs
)
# Align padding tokens between tokenizer and model
model.config.pad_token_id = tokenizer.pad_token_id
# If post-training a base model, use ChatML as the default template
if tokenizer.chat_template is None:
model, tokenizer = setup_chat_format(model, tokenizer)
if model_args.use_peft and model_args.lora_task_type != "SEQ_CLS":
warnings.warn(
"You are using a `task_type` that is different than `SEQ_CLS` for PEFT. This will lead to silent bugs"
" Make sure to pass --lora_task_type SEQ_CLS when using this script with PEFT.",
UserWarning,
)
##############
# Load dataset
##############
with open('./prefs.pkl', 'rb') as fh:
loaded_data = pickle.load(fh)
random.shuffle(loaded_data)
dataset = []
for a_wins, a, b in loaded_data:
if a_wins == 0:
a, b = b, a
dataset.append({'chosen': a, 'rejected': b})
dataset = Dataset.from_list(dataset)
# Split the dataset into training and evaluation sets
train_eval_split = dataset.train_test_split(test_size=0.15, shuffle=True, seed=42)
# Access the training and evaluation datasets
train_dataset = train_eval_split['train']
eval_dataset = train_eval_split['test']
##########
# Training
##########
trainer = RewardTrainer(
model=model,
processing_class=tokenizer,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
)
trainer.train()
```
Replacing `dataset = Dataset.from_list(dataset)` with
```python
with open('./prefs.json', 'w') as fh:
json.dump(dataset, fh)
dataset = load_dataset("json", data_files="./prefs.json", split='train')
```
resolves the issue.
| 2025-04-21T17:29:20 | 2025-06-29T06:20:45 | null |
https://github.com/huggingface/datasets/issues/7531
| null | 7,531 | false |
[
"Hi ! How big is the dataset ? if you load it using `from_list`, the dataset lives in memory and has to be copied to every gpu process, which can be slow.\n\nIt's fasted if you load it from JSON files from disk, because in that case the dataset in converted to Arrow and loaded from disk using memory mapping. Memory mapping allows to quickly reload the dataset in other processes.\n\nMaybe we can change `from_list` and other methods to always use the disk though, instead of loading in memory, WDYT ?",
"Thanks for raising this! As lhoestq mentioned, the root cause seems to be that `Dataset.from_list()` creates an in-memory dataset, which causes issues with DeepSpeed across multiple GPUs due to the cost of copying that memory to all processes.\n\nUsing `load_dataset(\"json\", ...)` works because Hugging Face datasets then convert the data to Apache Arrow and use **memory mapping**, which avoids this copying overhead.\n\nPossible improvement could be to add an option like `use_disk=True` to `Dataset.from_list()` to allow users to write to Arrow + memory-map the dataset, enabling compatibility with multi-process settings like DeepSpeed, while keeping the current fast behavior by default.\n\nWould love to hear if this direction sounds acceptable before attempting a PR.\n"
] |
3,007,452,499 |
How to solve "Spaces stuck in Building" problems
|
closed
|
### Describe the bug
Public spaces may stuck in Building after restarting, error log as follows:
build error
Unexpected job error
ERROR: failed to push spaces-registry.huggingface.tech/spaces/*:cpu-*-*: unexpected status from HEAD request to https://spaces-registry.huggingface.tech/v2/spaces/*/manifests/cpu-*-*: 401 Unauthorized
### Steps to reproduce the bug
Restart space / Factory rebuild cannot avoid it
### Expected behavior
Fix this problem
### Environment info
no requirements.txt can still happen
python gradio spaces
| 2025-04-21T03:08:38 | 2025-04-22T07:49:52 | 2025-04-22T07:49:52 |
https://github.com/huggingface/datasets/issues/7530
| null | 7,530 | false |
[
"I'm facing the same issue—Space stuck in \"Building\" even after restart and Factory rebuild. Any fix?\n",
"> I'm facing the same issue—Space stuck in \"Building\" even after restart and Factory rebuild. Any fix?\n\nAlso see https://github.com/huggingface/huggingface_hub/issues/3019",
"I'm facing the same issue. The build fails with the same error, and restarting won't help. Is there a fix or ETA? "
] |
3,007,118,969 |
audio folder builder cannot detect custom split name
|
open
|
### Describe the bug
when using audio folder builder (`load_dataset("audiofolder", data_dir="/path/to/folder")`), it cannot detect custom split name other than train/validation/test
### Steps to reproduce the bug
i have the following folder structure
```
my_dataset/
├── train/
│ ├── lorem.wav
│ ├── …
│ └── metadata.csv
├── test/
│ ├── ipsum.wav
│ ├── …
│ └── metadata.csv
├── validation/
│ ├── dolor.wav
│ ├── …
│ └── metadata.csv
└── custom/
├── sit.wav
├── …
└── metadata.csv
```
using `ds = load_dataset("audiofolder", data_dir="/path/to/my_dataset")`
### Expected behavior
i got `ds` with only 3 splits train/validation/test, whenever i rename train/validation/test folder it also disappear if i re-create `ds`
### Environment info
- `datasets` version: 3.5.0
- Platform: Windows-11-10.0.26100-SP0
- Python version: 3.12.8
- `huggingface_hub` version: 0.30.2
- PyArrow version: 18.1.0
- Pandas version: 2.2.3
- `fsspec` version: 2024.9.0
| 2025-04-20T16:53:21 | 2025-04-20T16:53:21 | null |
https://github.com/huggingface/datasets/issues/7529
| null | 7,529 | false |
[] |
3,006,433,485 |
Data Studio Error: Convert JSONL incorrectly
|
open
|
### Describe the bug
Hi there,
I uploaded a dataset here https://huggingface.co/datasets/V-STaR-Bench/V-STaR, but I found that Data Studio incorrectly convert the "bboxes" value for the whole dataset. Therefore, anyone who downloaded the dataset via the API would get the wrong "bboxes" value in the data file.
Could you help me address the issue?
Many thanks,
### Steps to reproduce the bug
The JSONL file of [V_STaR_test_release.jsonl](https://huggingface.co/datasets/V-STaR-Bench/V-STaR/blob/main/V_STaR_test_release.jsonl) has the correct values of every "bboxes" for each sample.
But in the Data Studio, we can see that the values of "bboxes" have changed, and load the dataset via API will also get the wrong values.
### Expected behavior
Fix the bug to correctly download my dataset.
### Environment info
- `datasets` version: 2.16.1
- Platform: Linux-5.14.0-427.22.1.el9_4.x86_64-x86_64-with-glibc2.34
- Python version: 3.10.16
- `huggingface_hub` version: 0.29.3
- PyArrow version: 19.0.0
- Pandas version: 2.2.3
- `fsspec` version: 2023.10.0
| 2025-04-19T13:21:44 | 2025-05-06T13:18:38 | null |
https://github.com/huggingface/datasets/issues/7528
| null | 7,528 | false |
[
"Hi ! Your JSONL file is incompatible with Arrow / Parquet. Indeed in Arrow / Parquet every dict should have the same keys, while in your dataset the bboxes have varying keys.\n\nThis causes the Data Studio to treat the bboxes as if each row was missing the keys from other rows.\n\nFeel free to take a look at the docs on object segmentation to see how to format a dataset with bboxes: https://huggingface.co/docs/datasets/object_detection"
] |
3,005,242,422 |
Auto-merge option for `convert-to-parquet`
|
closed
|
### Feature request
Add a command-line option, e.g. `--auto-merge-pull-request` that enables automatic merging of the commits created by the `convert-to-parquet` tool.
### Motivation
Large datasets may result in dozens of PRs due to the splitting mechanism. Each of these has to be manually accepted via the website.
### Your contribution
Happy to look into submitting a PR if this is of interest to maintainers.
| 2025-04-18T16:03:22 | 2025-07-18T19:09:03 | 2025-07-18T19:09:03 |
https://github.com/huggingface/datasets/issues/7527
| null | 7,527 | false |
[
"Alternatively, there could be an option to switch from submitting PRs to just committing changes directly to `main`.",
"Why not, I'd be in favor of `--merge-pull-request` to call `HfApi().merge_pull_request()` at the end of the conversion :) feel free to open a PR if you'd like",
"#self-assign",
"Closing since convert to parquet has been removed... https://github.com/huggingface/datasets/pull/7592#issuecomment-3073053138"
] |
3,005,107,536 |
Faster downloads/uploads with Xet storage
|
open
|
## Xet is out !
Over the past few weeks, Hugging Face’s [Xet Team](https://huggingface.co/xet-team) took a major step forward by [migrating the first Model and Dataset repositories off LFS and to Xet storage](https://huggingface.co/posts/jsulz/911431940353906).
See more information on the HF blog: https://huggingface.co/blog/xet-on-the-hub
You can already enable Xet on Hugging Face account to benefit from faster downloads and uploads :)
We finalized an official integration with the `huggingface_hub` library that means you get the benefits of Xet without any significant changes to your current workflow.
## Previous versions of `datasets`
For older versions of `datasets` you might see this warning in `push_to_hub()`:
```
Uploading files as bytes or binary IO objects is not supported by Xet Storage.
```
This means the `huggingface_hub` + Xet integration isn't enabled for your version of `datasets`.
You can fix this by updating to `datasets>=3.6.0` and `huggingface_hub>=0.31.0`
```
pip install -U datasets huggingface_hub
```
## The future
Stay tuned for more Xet optimizations, especially on [Xet-optimized Parquet](https://huggingface.co/blog/improve_parquet_dedupe)
| 2025-04-18T14:46:42 | 2025-05-12T12:09:09 | null |
https://github.com/huggingface/datasets/issues/7526
| null | 7,526 | false |
[] |
3,003,032,248 |
Fix indexing in split commit messages
|
closed
|
When a large commit is split up, it seems the commit index in the message is zero-based while the total number is one-based. I came across this running `convert-to-parquet` and was wondering why there was no `6-of-6` commit. This PR fixes that by adding one to the commit index, so both are one-based.
Current behavior:
<img width="463" alt="Screenshot 2025-04-17 at 1 00 17 PM" src="https://github.com/user-attachments/assets/7f3d389e-cb92-405d-a3c2-f2b1cdf0cb79" />
| 2025-04-17T17:06:26 | 2025-04-28T14:26:27 | 2025-04-28T14:26:27 |
https://github.com/huggingface/datasets/pull/7525
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7525",
"html_url": "https://github.com/huggingface/datasets/pull/7525",
"diff_url": "https://github.com/huggingface/datasets/pull/7525.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7525.patch",
"merged_at": null
}
| 7,525 | true |
[
"Hi ! this is expected and is coherent with other naming conventions in `datasets` such as parquet shards naming"
] |
3,002,067,826 |
correct use with polars example
|
closed
| null | 2025-04-17T10:19:19 | 2025-04-28T13:48:34 | 2025-04-28T13:48:33 |
https://github.com/huggingface/datasets/pull/7524
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7524",
"html_url": "https://github.com/huggingface/datasets/pull/7524",
"diff_url": "https://github.com/huggingface/datasets/pull/7524.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7524.patch",
"merged_at": "2025-04-28T13:48:33"
}
| 7,524 | true |
[] |
2,999,616,692 |
mention av in video docs
|
closed
| null | 2025-04-16T13:11:12 | 2025-04-16T13:13:45 | 2025-04-16T13:11:42 |
https://github.com/huggingface/datasets/pull/7523
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7523",
"html_url": "https://github.com/huggingface/datasets/pull/7523",
"diff_url": "https://github.com/huggingface/datasets/pull/7523.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7523.patch",
"merged_at": "2025-04-16T13:11:42"
}
| 7,523 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7523). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
2,998,169,017 |
Preserve formatting in concatenated IterableDataset
|
closed
|
Fixes #7515
| 2025-04-16T02:37:33 | 2025-05-19T15:07:38 | 2025-05-19T15:07:37 |
https://github.com/huggingface/datasets/pull/7522
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7522",
"html_url": "https://github.com/huggingface/datasets/pull/7522",
"diff_url": "https://github.com/huggingface/datasets/pull/7522.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7522.patch",
"merged_at": "2025-05-19T15:07:37"
}
| 7,522 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7522). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
2,997,666,366 |
fix: Image Feature in Datasets Library Fails to Handle bytearray Objects from Spark DataFrames (#7517)
|
closed
|
## Task
Support bytes-like objects (bytes and bytearray) in Features classes
### Description
The `Features` classes only accept `bytes` objects for binary data, but not `bytearray`. This leads to errors when using `IterableDataset.from_spark()` with Spark DataFrames as they contain `bytearray` objects, even though both `bytes` and `bytearray` are valid [*bytes-like objects* in Python](https://docs.python.org/3/glossary.html#term-bytes-like-object).
### Changes
- Updated `Features` classes to accept both `bytes` and `bytearray` types for binary data fields.
### Reasoning
- `bytes` and `bytearray` serve the same purpose for binary data, with the only difference being mutability.
- Modifying the Spark iterator to convert `bytearray` to `bytes` would be a workaround, not a true fix. I think the correct solution is to accept all bytes-like objects as input.
- This approach is more robust and future-proof since Python 3.12+ provides a [standard way to check for buffer protocol](https://docs.python.org/3/c-api/buffer.html#bufferobjects).
### Testing
- Added tests to cover `bytearray` inputs for image features.
### Related Issues
- Fixes: #7517
| 2025-04-15T21:23:58 | 2025-05-07T14:17:29 | 2025-05-07T14:17:29 |
https://github.com/huggingface/datasets/pull/7521
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7521",
"html_url": "https://github.com/huggingface/datasets/pull/7521",
"diff_url": "https://github.com/huggingface/datasets/pull/7521.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7521.patch",
"merged_at": "2025-05-07T14:17:29"
}
| 7,521 | true |
[
"@lhoestq let me know if you prefer to change the spark iterator so it outputs `bytes`",
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7521). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
2,997,422,044 |
Update items in the dataset without `map`
|
open
|
### Feature request
I would like to be able to update items in my dataset without affecting all rows. At least if there was a range option, I would be able to process those items, save the dataset, and then continue.
If I am supposed to split the dataset first, that is not clear, since the docs suggest that any of those functions returns a new object, so I don't think I can do that.
### Motivation
I am applying an extremely time-consuming function to each item in my `Dataset`. Unfortunately, datasets only supports updating values via `map`, so if my computer dies in the middle of this long-running process, I lose all progress. This is far from ideal. I would like to use `datasets` throughout this processing, but this limitation is now forcing me to write my own dataset format just to do this intermediary operation.
It would be less intuitive but I suppose I could split and then concatenate the dataset before saving? But this feels very inefficient.
### Your contribution
I can test the feature.
| 2025-04-15T19:39:01 | 2025-04-19T18:47:46 | null |
https://github.com/huggingface/datasets/issues/7520
| null | 7,520 | false |
[
"Hello!\n\nHave you looked at `Dataset.shard`? [Docs](https://huggingface.co/docs/datasets/en/process#shard)\n\nUsing this method you could break your dataset in N shards. Apply `map` on each shard and concatenate them back."
] |
2,996,458,961 |
pdf docs fixes
|
closed
|
close https://github.com/huggingface/datasets/issues/7494
| 2025-04-15T13:35:56 | 2025-04-15T13:38:31 | 2025-04-15T13:36:03 |
https://github.com/huggingface/datasets/pull/7519
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7519",
"html_url": "https://github.com/huggingface/datasets/pull/7519",
"diff_url": "https://github.com/huggingface/datasets/pull/7519.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7519.patch",
"merged_at": "2025-04-15T13:36:03"
}
| 7,519 | true |
[
"The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_7519). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update."
] |
2,996,141,825 |
num_proc parallelization works only for first ~10s.
|
open
|
### Describe the bug
When I try to load an already downloaded dataset with num_proc=64, the speed is very high for the first 10-20 seconds acheiving 30-40K samples / s, and 100% utilization for all cores but it soon drops to <= 1000 with almost 0% utilization for most cores.
### Steps to reproduce the bug
```
// download dataset with cli
!huggingface-cli download --repo-type dataset timm/imagenet-1k-wds --max-workers 32
from datasets import load_dataset
ds = load_dataset("timm/imagenet-1k-wds", num_proc=64)
```
### Expected behavior
100% core utilization throughout.
### Environment info
Azure A100-80GB, 16 cores VM
| 2025-04-15T11:44:03 | 2025-04-15T13:12:13 | null |
https://github.com/huggingface/datasets/issues/7518
| null | 7,518 | false |
[
"Hi, can you check if the processes are still alive ? It's a bit weird because `datasets` does check if processes crash and return an error in that case",
"Thank you for reverting quickly. I digged a bit, and realized my disk's IOPS is also limited - which is causing this. will check further and report if it's an issue of hf datasets' side or mine. "
] |
2,996,106,077 |
Image Feature in Datasets Library Fails to Handle bytearray Objects from Spark DataFrames
|
closed
|
### Describe the bug
When using `IterableDataset.from_spark()` with a Spark DataFrame containing image data, the `Image` feature class fails to properly process this data type, causing an `AttributeError: 'bytearray' object has no attribute 'get'`
### Steps to reproduce the bug
1. Create a Spark DataFrame with a column containing image data as bytearray objects
2. Define a Feature schema with an Image feature
3. Create an IterableDataset using `IterableDataset.from_spark()`
4. Attempt to iterate through the dataset
```
from pyspark.sql import SparkSession
from datasets import Dataset, IterableDataset, Features, Image, Value
# initialize spark
spark = SparkSession.builder.appName("MinimalRepro").getOrCreate()
# create spark dataframe
data = [(0, open("image.png", "rb").read())]
df = spark.createDataFrame(data, "idx: int, image: binary")
# convert to dataset
features = Features({"idx": Value("int64"), "image": Image()})
ds = Dataset.from_spark(df, features=features)
ds_iter = IterableDataset.from_spark(df, features=features)
# iterate
print(next(iter(ds)))
print(next(iter(ds_iter)))
```
### Expected behavior
The features should work on `IterableDataset` the same way they work on `Dataset`
### Environment info
- `datasets` version: 3.5.0
- Platform: macOS-15.3.2-arm64-arm-64bit
- Python version: 3.12.7
- `huggingface_hub` version: 0.30.2
- PyArrow version: 18.1.0
- Pandas version: 2.2.3
- `fsspec` version: 2024.12.0
| 2025-04-15T11:29:17 | 2025-05-07T14:17:30 | 2025-05-07T14:17:30 |
https://github.com/huggingface/datasets/issues/7517
| null | 7,517 | false |
[
"Hi ! The `Image()` type accepts either\n- a `bytes` object containing the image bytes\n- a `str` object containing the image path\n- a `PIL.Image` object\n\nbut it doesn't support `bytearray`, maybe you can convert to `bytes` beforehand ?",
"Hi @lhoestq, \nconverting to bytes is certainly possible and would work around the error. However, the core issue is that `Dataset` and `IterableDataset` behave differently with the features.\n\nI’d be happy to work on a fix for this issue.",
"I see, that's an issue indeed. Feel free to ping me if I can help with reviews or any guidance\n\nIf it can help, the code that takes a Spark DataFrame and iterates on the rows for `IterableDataset` is here: \n\nhttps://github.com/huggingface/datasets/blob/6a96bf313085d7538a999b929a550e14e1d406c9/src/datasets/packaged_modules/spark/spark.py#L49-L53",
"#self-assign"
] |
2,995,780,283 |
unsloth/DeepSeek-R1-Distill-Qwen-32B server error
|
closed
|
### Describe the bug
hfhubhttperror: 500 server error: internal server error for url: https://huggingface.co/api/models/unsloth/deepseek-r1-distill-qwen-32b-bnb-4bit/commits/main (request id: root=1-67fe23fa-3a2150eb444c2a823c388579;de3aed68-c397-4da5-94d4-6565efd3b919) internal error - we're working hard to fix this as soon as possible!
### Steps to reproduce the bug
unsloth/DeepSeek-R1-Distill-Qwen-32B server error
### Expected behavior
Network repair
### Environment info
The web side is also unavailable
| 2025-04-15T09:26:53 | 2025-04-15T09:57:26 | 2025-04-15T09:57:26 |
https://github.com/huggingface/datasets/issues/7516
| null | 7,516 | false |
[] |
2,995,082,418 |
`concatenate_datasets` does not preserve Pytorch format for IterableDataset
|
closed
|
### Describe the bug
When concatenating datasets with `concatenate_datasets`, I would expect the resulting combined dataset to be in the same format as the inputs (assuming it's consistent). This is indeed the behavior when combining `Dataset`, but not when combining `IterableDataset`. Specifically, when applying `concatenate_datasets` to a list of `IterableDataset` in Pytorch format (i.e. using `.with_format(Pytorch)`), the output `IterableDataset` is not in Pytorch format.
### Steps to reproduce the bug
```
import datasets
ds = datasets.Dataset.from_dict({"a": [1,2,3]})
iterable_ds = ds.to_iterable_dataset()
datasets.concatenate_datasets([ds.with_format("torch")]) # <- this preserves Pytorch format
datasets.concatenate_datasets([iterable_ds.with_format("torch")]) # <- this does NOT preserves Pytorch format
```
### Expected behavior
Pytorch format should be preserved when combining IterableDataset in Pytorch format.
### Environment info
datasets==3.5.0, Python 3.11.11, torch==2.2.2
| 2025-04-15T04:36:34 | 2025-05-19T15:07:38 | 2025-05-19T15:07:38 |
https://github.com/huggingface/datasets/issues/7515
| null | 7,515 | false |
[
"Hi ! Oh indeed it would be cool to return the same format in that case. Would you like to submit a PR ? The function that does the concatenation is here:\n\nhttps://github.com/huggingface/datasets/blob/90e5bf8a8599b625d6103ee5ac83b98269991141/src/datasets/iterable_dataset.py#L3375-L3380",
"Thank you for the pointer, @lhoestq ! See #7522 "
] |
2,994,714,923 |
Do not hash `generator` in `BuilderConfig.create_config_id`
|
closed
|
`Dataset.from_generator` function passes all of its arguments to `BuilderConfig.create_config_id`, including generator function itself. `BuilderConfig.create_config_id` function tries to hash all the args, and hashing a `generator` can take a large amount of time or even cause MemoryError if the dataset processed in a generator function is large enough.
Maybe we should pop generator from `config_kwargs_to_add_to_suffix` before hashing to avoid it.
There is a more detailed description of the problem this PR solves in #7513
| 2025-04-15T01:26:43 | 2025-04-23T11:55:55 | 2025-04-15T16:27:51 |
https://github.com/huggingface/datasets/pull/7514
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7514",
"html_url": "https://github.com/huggingface/datasets/pull/7514",
"diff_url": "https://github.com/huggingface/datasets/pull/7514.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7514.patch",
"merged_at": null
}
| 7,514 | true |
[] |
2,994,678,437 |
MemoryError while creating dataset from generator
|
open
|
### Describe the bug
# TL:DR
`Dataset.from_generator` function passes all of its arguments to `BuilderConfig.create_config_id`, including `generator` function itself. `BuilderConfig.create_config_id` function tries to hash all the args, which can take a large amount of time or even cause MemoryError if the dataset processed in a generator function is large enough.
Maybe we should pop `generator` from `config_kwargs_to_add_to_suffix` before hashing to avoid it.
# Full description
I have a pretty large spatial imagery dataset that is generated from two xbatcher.BatchGenerators via custom `dataset_generator` function that looks like this if simplified:
```
def dataset_generator():
for index in samples:
data_dict = {
"key": index,
"x": x_batches[index].data,
"y": y_batches[index].data,
}
yield data_dict
```
Then I use `datasets.Dataset.from_generator` to generate the dataset itself.
```
# Create dataset
ds = datasets.Dataset.from_generator(
dataset_generator,
features=feat,
cache_dir=(output / ".cache"),
)
```
It works nicely with pretty small data, but if the dataset is huge and barely fits in memory, it crashes with memory error:
<details>
<summary>Full stack trace</summary>
```
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\remote_sensing_processor\segmentation\semantic\tiles.py:248](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/remote_sensing_processor/segmentation/semantic/tiles.py#line=247), in generate_tiles(x, y, output, tile_size, shuffle, split, x_dtype, y_dtype, x_nodata, y_nodata)
245 yield data_dict
247 # Create dataset
--> 248 ds = datasets.Dataset.from_generator(
249 dataset_generator,
250 features=feat,
251 cache_dir=(output / ".cache"),
252 )
254 # Save dataset
255 ds.save_to_disk(output / name)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\arrow_dataset.py:1105](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/arrow_dataset.py#line=1104), in Dataset.from_generator(generator, features, cache_dir, keep_in_memory, gen_kwargs, num_proc, split, **kwargs)
1052 """Create a Dataset from a generator.
1053
1054 Args:
(...) 1101 ```
1102 """
1103 from .io.generator import GeneratorDatasetInputStream
-> 1105 return GeneratorDatasetInputStream(
1106 generator=generator,
1107 features=features,
1108 cache_dir=cache_dir,
1109 keep_in_memory=keep_in_memory,
1110 gen_kwargs=gen_kwargs,
1111 num_proc=num_proc,
1112 split=split,
1113 **kwargs,
1114 ).read()
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\io\generator.py:29](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/io/generator.py#line=28), in GeneratorDatasetInputStream.__init__(self, generator, features, cache_dir, keep_in_memory, streaming, gen_kwargs, num_proc, split, **kwargs)
9 def __init__(
10 self,
11 generator: Callable,
(...) 19 **kwargs,
20 ):
21 super().__init__(
22 features=features,
23 cache_dir=cache_dir,
(...) 27 **kwargs,
28 )
---> 29 self.builder = Generator(
30 cache_dir=cache_dir,
31 features=features,
32 generator=generator,
33 gen_kwargs=gen_kwargs,
34 split=split,
35 **kwargs,
36 )
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\builder.py:343](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/builder.py#line=342), in DatasetBuilder.__init__(self, cache_dir, dataset_name, config_name, hash, base_path, info, features, token, repo_id, data_files, data_dir, storage_options, writer_batch_size, **config_kwargs)
341 config_kwargs["data_dir"] = data_dir
342 self.config_kwargs = config_kwargs
--> 343 self.config, self.config_id = self._create_builder_config(
344 config_name=config_name,
345 custom_features=features,
346 **config_kwargs,
347 )
349 # prepare info: DatasetInfo are a standardized dataclass across all datasets
350 # Prefill datasetinfo
351 if info is None:
352 # TODO FOR PACKAGED MODULES IT IMPORTS DATA FROM src/packaged_modules which doesn't make sense
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\builder.py:604](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/builder.py#line=603), in DatasetBuilder._create_builder_config(self, config_name, custom_features, **config_kwargs)
598 builder_config._resolve_data_files(
599 base_path=self.base_path,
600 download_config=DownloadConfig(token=self.token, storage_options=self.storage_options),
601 )
603 # compute the config id that is going to be used for caching
--> 604 config_id = builder_config.create_config_id(
605 config_kwargs,
606 custom_features=custom_features,
607 )
608 is_custom = (config_id not in self.builder_configs) and config_id != "default"
609 if is_custom:
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\builder.py:187](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/builder.py#line=186), in BuilderConfig.create_config_id(self, config_kwargs, custom_features)
185 suffix = Hasher.hash(config_kwargs_to_add_to_suffix)
186 else:
--> 187 suffix = Hasher.hash(config_kwargs_to_add_to_suffix)
189 if custom_features is not None:
190 m = Hasher()
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\fingerprint.py:188](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/fingerprint.py#line=187), in Hasher.hash(cls, value)
186 @classmethod
187 def hash(cls, value: Any) -> str:
--> 188 return cls.hash_bytes(dumps(value))
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:109](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=108), in dumps(obj)
107 """Pickle an object to a string."""
108 file = BytesIO()
--> 109 dump(obj, file)
110 return file.getvalue()
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:103](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=102), in dump(obj, file)
101 def dump(obj, file):
102 """Pickle an object to a file."""
--> 103 Pickler(file, recurse=True).dump(obj)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:420](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=419), in Pickler.dump(self, obj)
418 def dump(self, obj): #NOTE: if settings change, need to update attributes
419 logger.trace_setup(self)
--> 420 StockPickler.dump(self, obj)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:484](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=483), in _Pickler.dump(self, obj)
482 if self.proto >= 4:
483 self.framer.start_framing()
--> 484 self.save(obj)
485 self.write(STOP)
486 self.framer.end_framing()
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:1217](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=1216), in save_module_dict(pickler, obj)
1214 if is_dill(pickler, child=False) and pickler._session:
1215 # we only care about session the first pass thru
1216 pickler._first_pass = False
-> 1217 StockPickler.save_dict(pickler, obj)
1218 logger.trace(pickler, "# D2")
1219 return
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:990](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=989), in _Pickler.save_dict(self, obj)
987 self.write(MARK + DICT)
989 self.memoize(obj)
--> 990 self._batch_setitems(obj.items())
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:83](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=82), in Pickler._batch_setitems(self, items)
80 from datasets.fingerprint import Hasher
82 items = sorted(items, key=lambda x: Hasher.hash(x[0]))
---> 83 dill.Pickler._batch_setitems(self, items)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:1014](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=1013), in _Pickler._batch_setitems(self, items)
1012 for k, v in tmp:
1013 save(k)
-> 1014 save(v)
1015 write(SETITEMS)
1016 elif n:
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:1985](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=1984), in save_function(pickler, obj)
1982 if state_dict:
1983 state = state, state_dict
-> 1985 _save_with_postproc(pickler, (_create_function, (
1986 obj.__code__, globs, obj.__name__, obj.__defaults__,
1987 closure
1988 ), state), obj=obj, postproc_list=postproc_list)
1990 # Lift closure cell update to earliest function (#458)
1991 if _postproc:
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:1117](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=1116), in _save_with_postproc(pickler, reduction, is_pickler_dill, obj, postproc_list)
1115 continue
1116 else:
-> 1117 pickler.save_reduce(*reduction)
1118 # pop None created by calling preprocessing step off stack
1119 pickler.write(POP)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:690](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=689), in _Pickler.save_reduce(self, func, args, state, listitems, dictitems, state_setter, obj)
688 else:
689 save(func)
--> 690 save(args)
691 write(REDUCE)
693 if obj is not None:
694 # If the object is already in the memo, this means it is
695 # recursive. In this case, throw away everything we put on the
696 # stack, and fetch the object back from the memo.
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:905](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=904), in _Pickler.save_tuple(self, obj)
903 if n <= 3 and self.proto >= 2:
904 for element in obj:
--> 905 save(element)
906 # Subtle. Same as in the big comment below.
907 if id(obj) in memo:
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:601](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=600), in _Pickler.save(self, obj, save_persistent_id)
597 raise PicklingError("Tuple returned by %s must have "
598 "two to six elements" % reduce)
600 # Save the reduce() output and finally memoize the object
--> 601 self.save_reduce(obj=obj, *rv)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:715](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=714), in _Pickler.save_reduce(self, func, args, state, listitems, dictitems, state_setter, obj)
713 if state is not None:
714 if state_setter is None:
--> 715 save(state)
716 write(BUILD)
717 else:
718 # If a state_setter is specified, call it instead of load_build
719 # to update obj's with its previous state.
720 # First, push state_setter and its tuple of expected arguments
721 # (obj, state) onto the stack.
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:1217](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=1216), in save_module_dict(pickler, obj)
1214 if is_dill(pickler, child=False) and pickler._session:
1215 # we only care about session the first pass thru
1216 pickler._first_pass = False
-> 1217 StockPickler.save_dict(pickler, obj)
1218 logger.trace(pickler, "# D2")
1219 return
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:990](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=989), in _Pickler.save_dict(self, obj)
987 self.write(MARK + DICT)
989 self.memoize(obj)
--> 990 self._batch_setitems(obj.items())
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:83](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=82), in Pickler._batch_setitems(self, items)
80 from datasets.fingerprint import Hasher
82 items = sorted(items, key=lambda x: Hasher.hash(x[0]))
---> 83 dill.Pickler._batch_setitems(self, items)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:1014](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=1013), in _Pickler._batch_setitems(self, items)
1012 for k, v in tmp:
1013 save(k)
-> 1014 save(v)
1015 write(SETITEMS)
1016 elif n:
[... skipping similar frames: Pickler.save at line 70 (1 times), Pickler.save at line 414 (1 times)]
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:601](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=600), in _Pickler.save(self, obj, save_persistent_id)
597 raise PicklingError("Tuple returned by %s must have "
598 "two to six elements" % reduce)
600 # Save the reduce() output and finally memoize the object
--> 601 self.save_reduce(obj=obj, *rv)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:715](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=714), in _Pickler.save_reduce(self, func, args, state, listitems, dictitems, state_setter, obj)
713 if state is not None:
714 if state_setter is None:
--> 715 save(state)
716 write(BUILD)
717 else:
718 # If a state_setter is specified, call it instead of load_build
719 # to update obj's with its previous state.
720 # First, push state_setter and its tuple of expected arguments
721 # (obj, state) onto the stack.
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:905](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=904), in _Pickler.save_tuple(self, obj)
903 if n <= 3 and self.proto >= 2:
904 for element in obj:
--> 905 save(element)
906 # Subtle. Same as in the big comment below.
907 if id(obj) in memo:
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:1217](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=1216), in save_module_dict(pickler, obj)
1214 if is_dill(pickler, child=False) and pickler._session:
1215 # we only care about session the first pass thru
1216 pickler._first_pass = False
-> 1217 StockPickler.save_dict(pickler, obj)
1218 logger.trace(pickler, "# D2")
1219 return
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:990](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=989), in _Pickler.save_dict(self, obj)
987 self.write(MARK + DICT)
989 self.memoize(obj)
--> 990 self._batch_setitems(obj.items())
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:83](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=82), in Pickler._batch_setitems(self, items)
80 from datasets.fingerprint import Hasher
82 items = sorted(items, key=lambda x: Hasher.hash(x[0]))
---> 83 dill.Pickler._batch_setitems(self, items)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:1014](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=1013), in _Pickler._batch_setitems(self, items)
1012 for k, v in tmp:
1013 save(k)
-> 1014 save(v)
1015 write(SETITEMS)
1016 elif n:
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:601](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=600), in _Pickler.save(self, obj, save_persistent_id)
597 raise PicklingError("Tuple returned by %s must have "
598 "two to six elements" % reduce)
600 # Save the reduce() output and finally memoize the object
--> 601 self.save_reduce(obj=obj, *rv)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:715](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=714), in _Pickler.save_reduce(self, func, args, state, listitems, dictitems, state_setter, obj)
713 if state is not None:
714 if state_setter is None:
--> 715 save(state)
716 write(BUILD)
717 else:
718 # If a state_setter is specified, call it instead of load_build
719 # to update obj's with its previous state.
720 # First, push state_setter and its tuple of expected arguments
721 # (obj, state) onto the stack.
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:905](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=904), in _Pickler.save_tuple(self, obj)
903 if n <= 3 and self.proto >= 2:
904 for element in obj:
--> 905 save(element)
906 # Subtle. Same as in the big comment below.
907 if id(obj) in memo:
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:1217](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=1216), in save_module_dict(pickler, obj)
1214 if is_dill(pickler, child=False) and pickler._session:
1215 # we only care about session the first pass thru
1216 pickler._first_pass = False
-> 1217 StockPickler.save_dict(pickler, obj)
1218 logger.trace(pickler, "# D2")
1219 return
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:990](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=989), in _Pickler.save_dict(self, obj)
987 self.write(MARK + DICT)
989 self.memoize(obj)
--> 990 self._batch_setitems(obj.items())
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:83](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=82), in Pickler._batch_setitems(self, items)
80 from datasets.fingerprint import Hasher
82 items = sorted(items, key=lambda x: Hasher.hash(x[0]))
---> 83 dill.Pickler._batch_setitems(self, items)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:1014](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=1013), in _Pickler._batch_setitems(self, items)
1012 for k, v in tmp:
1013 save(k)
-> 1014 save(v)
1015 write(SETITEMS)
1016 elif n:
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:601](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=600), in _Pickler.save(self, obj, save_persistent_id)
597 raise PicklingError("Tuple returned by %s must have "
598 "two to six elements" % reduce)
600 # Save the reduce() output and finally memoize the object
--> 601 self.save_reduce(obj=obj, *rv)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:690](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=689), in _Pickler.save_reduce(self, func, args, state, listitems, dictitems, state_setter, obj)
688 else:
689 save(func)
--> 690 save(args)
691 write(REDUCE)
693 if obj is not None:
694 # If the object is already in the memo, this means it is
695 # recursive. In this case, throw away everything we put on the
696 # stack, and fetch the object back from the memo.
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:920](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=919), in _Pickler.save_tuple(self, obj)
918 write(MARK)
919 for element in obj:
--> 920 save(element)
922 if id(obj) in memo:
923 # Subtle. d was not in memo when we entered save_tuple(), so
924 # the process of saving the tuple's elements must have saved
(...) 928 # could have been done in the "for element" loop instead, but
929 # recursive tuples are a rare thing.
930 get = self.get(memo[id(obj)][0])
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:601](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=600), in _Pickler.save(self, obj, save_persistent_id)
597 raise PicklingError("Tuple returned by %s must have "
598 "two to six elements" % reduce)
600 # Save the reduce() output and finally memoize the object
--> 601 self.save_reduce(obj=obj, *rv)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:715](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=714), in _Pickler.save_reduce(self, func, args, state, listitems, dictitems, state_setter, obj)
713 if state is not None:
714 if state_setter is None:
--> 715 save(state)
716 write(BUILD)
717 else:
718 # If a state_setter is specified, call it instead of load_build
719 # to update obj's with its previous state.
720 # First, push state_setter and its tuple of expected arguments
721 # (obj, state) onto the stack.
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:1217](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=1216), in save_module_dict(pickler, obj)
1214 if is_dill(pickler, child=False) and pickler._session:
1215 # we only care about session the first pass thru
1216 pickler._first_pass = False
-> 1217 StockPickler.save_dict(pickler, obj)
1218 logger.trace(pickler, "# D2")
1219 return
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:990](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=989), in _Pickler.save_dict(self, obj)
987 self.write(MARK + DICT)
989 self.memoize(obj)
--> 990 self._batch_setitems(obj.items())
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:83](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=82), in Pickler._batch_setitems(self, items)
80 from datasets.fingerprint import Hasher
82 items = sorted(items, key=lambda x: Hasher.hash(x[0]))
---> 83 dill.Pickler._batch_setitems(self, items)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:1014](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=1013), in _Pickler._batch_setitems(self, items)
1012 for k, v in tmp:
1013 save(k)
-> 1014 save(v)
1015 write(SETITEMS)
1016 elif n:
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:1217](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=1216), in save_module_dict(pickler, obj)
1214 if is_dill(pickler, child=False) and pickler._session:
1215 # we only care about session the first pass thru
1216 pickler._first_pass = False
-> 1217 StockPickler.save_dict(pickler, obj)
1218 logger.trace(pickler, "# D2")
1219 return
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:990](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=989), in _Pickler.save_dict(self, obj)
987 self.write(MARK + DICT)
989 self.memoize(obj)
--> 990 self._batch_setitems(obj.items())
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:83](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=82), in Pickler._batch_setitems(self, items)
80 from datasets.fingerprint import Hasher
82 items = sorted(items, key=lambda x: Hasher.hash(x[0]))
---> 83 dill.Pickler._batch_setitems(self, items)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:1019](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=1018), in _Pickler._batch_setitems(self, items)
1017 k, v = tmp[0]
1018 save(k)
-> 1019 save(v)
1020 write(SETITEM)
1021 # else tmp is empty, and we're done
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:601](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=600), in _Pickler.save(self, obj, save_persistent_id)
597 raise PicklingError("Tuple returned by %s must have "
598 "two to six elements" % reduce)
600 # Save the reduce() output and finally memoize the object
--> 601 self.save_reduce(obj=obj, *rv)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:715](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=714), in _Pickler.save_reduce(self, func, args, state, listitems, dictitems, state_setter, obj)
713 if state is not None:
714 if state_setter is None:
--> 715 save(state)
716 write(BUILD)
717 else:
718 # If a state_setter is specified, call it instead of load_build
719 # to update obj's with its previous state.
720 # First, push state_setter and its tuple of expected arguments
721 # (obj, state) onto the stack.
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:1217](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=1216), in save_module_dict(pickler, obj)
1214 if is_dill(pickler, child=False) and pickler._session:
1215 # we only care about session the first pass thru
1216 pickler._first_pass = False
-> 1217 StockPickler.save_dict(pickler, obj)
1218 logger.trace(pickler, "# D2")
1219 return
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:990](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=989), in _Pickler.save_dict(self, obj)
987 self.write(MARK + DICT)
989 self.memoize(obj)
--> 990 self._batch_setitems(obj.items())
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:83](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=82), in Pickler._batch_setitems(self, items)
80 from datasets.fingerprint import Hasher
82 items = sorted(items, key=lambda x: Hasher.hash(x[0]))
---> 83 dill.Pickler._batch_setitems(self, items)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:1014](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=1013), in _Pickler._batch_setitems(self, items)
1012 for k, v in tmp:
1013 save(k)
-> 1014 save(v)
1015 write(SETITEMS)
1016 elif n:
[... skipping similar frames: Pickler.save at line 70 (1 times), Pickler.save at line 414 (1 times)]
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:1217](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=1216), in save_module_dict(pickler, obj)
1214 if is_dill(pickler, child=False) and pickler._session:
1215 # we only care about session the first pass thru
1216 pickler._first_pass = False
-> 1217 StockPickler.save_dict(pickler, obj)
1218 logger.trace(pickler, "# D2")
1219 return
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:990](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=989), in _Pickler.save_dict(self, obj)
987 self.write(MARK + DICT)
989 self.memoize(obj)
--> 990 self._batch_setitems(obj.items())
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:83](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=82), in Pickler._batch_setitems(self, items)
80 from datasets.fingerprint import Hasher
82 items = sorted(items, key=lambda x: Hasher.hash(x[0]))
---> 83 dill.Pickler._batch_setitems(self, items)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:1014](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=1013), in _Pickler._batch_setitems(self, items)
1012 for k, v in tmp:
1013 save(k)
-> 1014 save(v)
1015 write(SETITEMS)
1016 elif n:
[... skipping similar frames: Pickler.save at line 70 (1 times), Pickler.save at line 414 (1 times)]
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:601](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=600), in _Pickler.save(self, obj, save_persistent_id)
597 raise PicklingError("Tuple returned by %s must have "
598 "two to six elements" % reduce)
600 # Save the reduce() output and finally memoize the object
--> 601 self.save_reduce(obj=obj, *rv)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:715](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=714), in _Pickler.save_reduce(self, func, args, state, listitems, dictitems, state_setter, obj)
713 if state is not None:
714 if state_setter is None:
--> 715 save(state)
716 write(BUILD)
717 else:
718 # If a state_setter is specified, call it instead of load_build
719 # to update obj's with its previous state.
720 # First, push state_setter and its tuple of expected arguments
721 # (obj, state) onto the stack.
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:920](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=919), in _Pickler.save_tuple(self, obj)
918 write(MARK)
919 for element in obj:
--> 920 save(element)
922 if id(obj) in memo:
923 # Subtle. d was not in memo when we entered save_tuple(), so
924 # the process of saving the tuple's elements must have saved
(...) 928 # could have been done in the "for element" loop instead, but
929 # recursive tuples are a rare thing.
930 get = self.get(memo[id(obj)][0])
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\datasets\utils\_dill.py:70](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/datasets/utils/_dill.py#line=69), in Pickler.save(self, obj, save_persistent_id)
68 if obj_type is FunctionType:
69 obj = getattr(obj, "_torchdynamo_orig_callable", obj)
---> 70 dill.Pickler.save(self, obj, save_persistent_id=save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\site-packages\dill\_dill.py:414](file:///C:/ProgramData/miniforge3/envs/geo/Lib/site-packages/dill/_dill.py#line=413), in Pickler.save(self, obj, save_persistent_id)
412 msg = "Can't pickle %s: attribute lookup builtins.generator failed" % GeneratorType
413 raise PicklingError(msg)
--> 414 StockPickler.save(self, obj, save_persistent_id)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:558](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=557), in _Pickler.save(self, obj, save_persistent_id)
556 f = self.dispatch.get(t)
557 if f is not None:
--> 558 f(self, obj) # Call unbound method with explicit self
559 return
561 # Check private dispatch table if any, or else
562 # copyreg.dispatch_table
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:809](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=808), in _Pickler.save_bytes(self, obj)
806 self.save_reduce(codecs.encode,
807 (str(obj, 'latin1'), 'latin1'), obj=obj)
808 return
--> 809 self._save_bytes_no_memo(obj)
810 self.memoize(obj)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:797](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=796), in _Pickler._save_bytes_no_memo(self, obj)
795 self._write_large_bytes(BINBYTES8 + pack("<Q", n), obj)
796 elif n >= self.framer._FRAME_SIZE_TARGET:
--> 797 self._write_large_bytes(BINBYTES + pack("<I", n), obj)
798 else:
799 self.write(BINBYTES + pack("<I", n) + obj)
File [C:\ProgramData\miniforge3\envs\geo\Lib\pickle.py:254](file:///C:/ProgramData/miniforge3/envs/geo/Lib/pickle.py#line=253), in _Framer.write_large_bytes(self, header, payload)
247 # Perform direct write of the header and payload of the large binary
248 # object. Be careful not to concatenate the header and the payload
249 # prior to calling 'write' as we do not want to allocate a large
250 # temporary bytes object.
251 # We intentionally do not insert a protocol 4 frame opcode to make
252 # it possible to optimize file.read calls in the loader.
253 write(header)
--> 254 write(payload)
MemoryError:
```
</details>
Memory error is an expected type of error in such case, but when I started digging down, I found out that it occurs in a kinda unexpected place - in `create_config_id` function. It tries to hash `config_kwargs_to_add_to_suffix`, including generator function itself.
I modified the `BuilderConfig.create_config_id` code like this to check which values are hashed and how much time it takes to hash them and ran it on a toy dataset:
```
print(config_kwargs_to_add_to_suffix)
start_time = time.time()
if all(isinstance(v, (str, bool, int, float)) for v in config_kwargs_to_add_to_suffix.values()):
suffix = ",".join(
str(k) + "=" + urllib.parse.quote_plus(str(v)) for k, v in config_kwargs_to_add_to_suffix.items()
)
if len(suffix) > 32: # hash if too long
suffix = Hasher.hash(config_kwargs_to_add_to_suffix)
else:
suffix = Hasher.hash(config_kwargs_to_add_to_suffix)
end_time = time.time()
print(f"Execution time: {end_time - start_time:.4f} seconds")
print(suffix)
```
In my case the content of `config_kwargs_to_add_to_suffix` was like this:
```
{'features': {'key': Value(dtype='int64', id=None), 'x': Array3D(shape=(44, 128, 128), dtype='float32', id=None), 'y_class': Array2D(shape=(128, 128), dtype='int32', id=None)}, 'gen_kwargs': None, 'generator': <function generate_tiles.<locals>.dataset_generator at 0x00000139D10D7920>, 'split': NamedSplit('train')}
```
Also I noticed that hashing took a significant amount of time - 43.1482 seconds, while the overall function execution (with data loading, batching and saving dataset) took 2min 45s. The output of `create_config_id` is just a dataset id, so, it is inappropirately large amount of time.
But when I added `config_kwargs_to_add_to_suffix.pop("generator", None)`, the hashing took only 0.0060 seconds.
Maybe we shouldn't hash the generator function, as it can be really computationally and memory expensive.
### Steps to reproduce the bug
This is a simplified example of a workflow I used to generate dataset. But I think that you can use almost any workflow to reproduce that bug.
```
import pystac
import pystac_client
import planetary_computer
import numpy as np
import xarray as xr
import rioxarray as rxr
import dask
import xbatcher
import datasets
# Loading a dataset, in our case - single Landsat image
catalog = pystac_client.Client.open(
"https://planetarycomputer.microsoft.com/api/stac/v1",
modifier=planetary_computer.sign_inplace,
)
brazil = [-60.2, -3.31]
time_of_interest = "2021-06-01/2021-08-31"
search = catalog.search(collections=["landsat-c2-l2"], intersects={"type": "Point", "coordinates": brazil}, datetime=time_of_interest)
items = search.item_collection()
item = min(items, key=lambda item: pystac.extensions.eo.EOExtension.ext(item).cloud_cover)
# Getting x data
bands = []
for band in ["red", "green", "blue", "nir08", "coastal", "swir16", "swir22", "lwir11"]:
with rxr.open_rasterio(item.assets[band].href, chunks=True, lock=True) as raster:
raster = raster.to_dataset('band')
#print(raster)
raster = raster.rename({1: band})
bands.append(raster)
x = xr.merge(bands).squeeze().to_array("band").persist()
# Getting y data
with rxr.open_rasterio(item.assets['qa_pixel'].href, chunks=True, lock=True) as raster:
y = raster.squeeze().persist()
# Setting up batches generators
x_batches = xbatcher.BatchGenerator(ds=x, input_dims={"x": 256, "y": 256})
y_batches = xbatcher.BatchGenerator(ds=y, input_dims={"x": 256, "y": 256})
# Filtering samples that contain only nodata
samples = list(range(len(x_batches)))
samples_filtered = []
for i in samples:
if not np.array_equal(np.unique(x_batches[i]), np.array([0.])) and not np.array_equal(np.unique(y_batches[i]), np.array([0])):
samples_filtered.append(i)
samples = samples_filtered
np.random.shuffle(samples)
# Setting up features
feat = {
"key": datasets.Value(dtype="int64"),
"x": datasets.Array3D(dtype="float32", shape=(4, 256, 256)),
"y": datasets.Array2D(dtype="int32", shape=(256, 256))
}
feat = datasets.Features(feat)
# Setting up a generator
def dataset_generator():
for index in samples:
data_dict = {
"key": index,
"x": x_batches[index].data,
"y": y_batches[index].data,
}
yield data_dict
# Create dataset
ds = datasets.Dataset.from_generator(
dataset_generator,
features=feat,
cache_dir="temp/cache",
)
```
Please, try adding `config_kwargs_to_add_to_suffix.pop("generator", None)` to `BuilderConfig.create_config_id` and then measuring how much time it takes to run
```
if all(isinstance(v, (str, bool, int, float)) for v in config_kwargs_to_add_to_suffix.values()):
suffix = ",".join(
str(k) + "=" + urllib.parse.quote_plus(str(v)) for k, v in config_kwargs_to_add_to_suffix.items()
)
if len(suffix) > 32: # hash if too long
suffix = Hasher.hash(config_kwargs_to_add_to_suffix)
else:
suffix = Hasher.hash(config_kwargs_to_add_to_suffix)
```
code block with and without `config_kwargs_to_add_to_suffix.pop("generator", None)`
In my case the difference was 3.3828 seconds without popping generator function and 0.0010 seconds with popping.
### Expected behavior
Much faster hashing and no MemoryErrors
### Environment info
- `datasets` version: 3.5.0
- Platform: Windows-11-10.0.26100-SP0
- Python version: 3.12.9
- `huggingface_hub` version: 0.30.1
- PyArrow version: 17.0.0
- Pandas version: 2.2.2
- `fsspec` version: 2024.12.0
| 2025-04-15T01:02:02 | 2025-04-23T19:37:08 | null |
https://github.com/huggingface/datasets/issues/7513
| null | 7,513 | false |
[
"Upd: created a PR that can probably solve the problem: #7514",
"Hi ! We need to take the generator into account for the cache. The generator is hashed to make the dataset fingerprint used by the cache. This way you can reload the Dataset from the cache without regenerating in subsequent `from_generator` calls.\n\nMaybe instead of removing generator from the hasher input, we can let users pass their own Dataset fingerprint to `from_generator`, and if it's specified we don't need to hash anything",
"Upd: I successfully generated a dataset from my large geospatial data with `generator` excluded from hashing and saved it to disk without running into memory errors. So, it looks like there are no other bottlenecks in dataset generation in my case\n\nMaybe letting users pass their own fingerprint to skip hashing can be a great solution to that issue!",
"@lhoestq I tried to implement user-defined dataset fingerprint in #7533 . Am I doing it right?"
] |
2,994,043,544 |
.map() fails if function uses pyvista
|
open
|
### Describe the bug
Using PyVista inside a .map() produces a crash with `objc[78796]: +[NSResponder initialize] may have been in progress in another thread when fork() was called. We cannot safely call it or ignore it in the fork() child process. Crashing instead. Set a breakpoint on objc_initializeAfterForkError to debug.`
### Steps to reproduce the bug
Run
```python
import numpy as np
import pyvista as pv
import datasets
data = [{"coords": np.random.rand(5, 3)} for _ in range(3)]
def render_point(example):
plotter = pv.Plotter(off_screen=True)
cloud = pv.PolyData(example["coords"])
plotter.add_mesh(cloud)
img = plotter.screenshot(return_img=True)
return {"image": img}
# breaks if num_proc>1
ds = datasets.Dataset.from_list(data).map(render_point, num_proc=2)
```
### Expected behavior
It should work. Just like when I use a process pool to make it work.
```python
import numpy as np
import pyvista as pv
import multiprocessing
data = [{"coords": np.random.rand(5, 3)} for _ in range(3)]
def render_point(example):
plotter = pv.Plotter(off_screen=True)
cloud = pv.PolyData(example["coords"])
plotter.add_mesh(cloud)
img = plotter.screenshot(return_img=True)
return {"image": img}
if __name__ == "__main__":
with multiprocessing.Pool(processes=2) as pool:
results = pool.map(render_point, data)
print(results[0]["image"].shape)
```
### Environment info
- `datasets` version: 3.3.2
- Platform: macOS-15.3.2-arm64-arm-64bit
- Python version: 3.11.10
- `huggingface_hub` version: 0.28.1
- PyArrow version: 18.1.0
- Pandas version: 2.2.3
- `fsspec` version: 2024.10.0
And then I suppose pyvista info is good to have.
```python
import pyvista as pv
print(pv.Report())
```
gives
--------------------------------------------------------------------------------
Date: Mon Apr 14 21:42:08 2025 CEST
OS : Darwin (macOS 15.3.2)
CPU(s) : 10
Machine : arm64
Architecture : 64bit
RAM : 32.0 GiB
Environment : IPython
File system : apfs
GPU Vendor : Apple
GPU Renderer : Apple M1 Max
GPU Version : 4.1 Metal - 89.3
MathText Support : True
Python 3.11.10 (main, Oct 7 2024, 23:25:27) [Clang 18.1.8 ]
pyvista : 0.44.2
vtk : 9.4.0
numpy : 2.2.2
matplotlib : 3.10.0
scooby : 0.10.0
pooch : 1.8.2
pillow : 11.1.0
imageio : 2.36.1
PyQt5 : 5.15.11
IPython : 8.30.0
scipy : 1.14.1
tqdm : 4.67.1
jupyterlab : 4.3.5
nest_asyncio : 1.6.0
--------------------------------------------------------------------------------
| 2025-04-14T19:43:02 | 2025-04-14T20:01:53 | null |
https://github.com/huggingface/datasets/issues/7512
| null | 7,512 | false |
[
"I found a similar (?) issue in https://github.com/huggingface/datasets/issues/6435, where someone had issues with forks and CUDA. According to https://huggingface.co/docs/datasets/main/en/process#multiprocessing we should do \n\n```\nfrom multiprocess import set_start_method\nset_start_method(\"spawn\")\n```\n\nto avoid the fork. The updated code\n\n```python\nimport numpy as np\nimport pyvista as pv\nimport datasets\nimport multiprocess\n\ndata = [{\"coords\": np.random.rand(5, 3)} for _ in range(3)]\n\ndef render_point(example):\n plotter = pv.Plotter(off_screen=True)\n cloud = pv.PolyData(example[\"coords\"])\n plotter.add_mesh(cloud)\n img = plotter.screenshot(return_img=True)\n return {\"image\": img}\n\n\n# breaks if num_proc>1\nmultiprocess.set_start_method(\"spawn\")\nds = datasets.Dataset.from_list(data).map(render_point, num_proc=2)\n```\n\ninstead fails with `TypeError: fork_exec() takes exactly 23 arguments (21 given)` which also seems like a bug to me."
] |
2,992,131,117 |
Incompatibile dill version (0.3.9) in datasets 2.18.0 - 3.5.0
|
open
|
### Describe the bug
Datasets 2.18.0 - 3.5.0 has a dependency on dill < 0.3.9. This causes errors with dill >= 0.3.9.
Could you please take a look into it and make it compatible?
### Steps to reproduce the bug
1. Install setuptools >= 2.18.0
2. Install dill >=0.3.9
3. Run pip check
4. Output:
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
datasets 2.18.0 requires dill<0.3.9,>=0.3.0, but you have dill 0.3.9 which is incompatible.
### Expected behavior
Pip install both libraries without any errors
### Environment info
Datasets version: 2.18 - 3.5
Python: 3.11
| 2025-04-14T07:22:44 | 2025-05-19T14:54:04 | null |
https://github.com/huggingface/datasets/issues/7510
| null | 7,510 | false |
[
"Hi ! We can bump `dill` to 0.3.9 if we make sure it's deterministic and doesn't break the caching mechanism in `datasets`.\n\nWould you be interested in opening a PR ? Then we can run the CI to see if it works",
"Hi!. Yeah I can do it. Should I make any changes besides dill versions?",
"There are probably some usage of internal functions from `dill` that we'll need to update in `datasets`\n\nIf you run `pytest tests/test_fingerprint.py` you should already have a good idea of what works and what doesn't.\nBut feel free to open a PR anyway, this way we can run the full CI and see the results\n",
"Hi, sorry for no response from my side. I will try to do it today.",
"Created pull request: [LINK](https://github.com/huggingface/datasets/pull/7535)\nTried to run tests by using command you have send and got few errors:\n\n",
"Thanks for running the test ! So it appears we have two issues to fix:\n1. 'log' is not defined: it seems an internal `dill` function has disappeared, so we should adapt the `datasets` code that was using it\n2. there are some hashes mismatches, which means `dill` doesn't seem to output the same dump when passed the same ipython function twice, or the same function but located at a different line in a python file"
] |
2,991,484,542 |
Dataset uses excessive memory when loading files
|
open
|
### Describe the bug
Hi
I am having an issue when loading a dataset.
I have about 200 json files each about 1GB (total about 215GB). each row has a few features which are a list of ints.
I am trying to load the dataset using `load_dataset`.
The dataset is about 1.5M samples
I use `num_proc=32` and a node with 378GB of memory.
About a third of the way there I get an OOM.
I also saw an old bug with a similar issue, which says to set `writer_batch_size`. I tried to lower it to 10, but it still crashed.
I also tried to lower the `num_proc` to 16 and even 8, but still the same issue.
### Steps to reproduce the bug
`dataset = load_dataset("json", data_dir=data_config.train_path, num_proc=data_config.num_proc, writer_batch_size=50)["train"]`
### Expected behavior
Loading a dataset with more than 100GB to spare should not cause an OOM error.
maybe i am missing something but I would love some help.
### Environment info
- `datasets` version: 3.5.0
- Platform: Linux-6.6.20-aufs-1-x86_64-with-glibc2.36
- Python version: 3.11.2
- `huggingface_hub` version: 0.29.1
- PyArrow version: 19.0.1
- Pandas version: 2.2.3
- `fsspec` version: 2024.9.0
| 2025-04-13T21:09:49 | 2025-04-28T15:18:55 | null |
https://github.com/huggingface/datasets/issues/7509
| null | 7,509 | false |
[
"small update: I converted the jsons to parquet and it now works well with 32 proc and the same node. \nI still think this needs to be understood, since json is a very popular and easy-to-use format. ",
"Hi ! The JSON loader loads full files in memory, unless they are JSON Lines. In this case it iterates on the JSON Lines in a memory efficient manner.\n\nI know there is an `ijson` package that works similarly but for general JSON files, maybe it can help and remove the need to load full JSON files in memory",
"Hi, i understand that json files are probably loaded into memory to read them but aren't they released when we write all the file content into arrow or something? ",
"Yes correct, the JSON data is only in memory during the conversion to Arrow. Then, the data is memory mapped from you disk",
"so the json files are all loaded into memory before converting to arrow? or do they convert 1 json at a time and then they are realeased?\nI don't understand how 200GB worth of jsons fill a 378GB node's memory.",
"Each process converts one JSON file at at time, So the total memory usage is num_proc * json_file_size * overhead, where overhead can be around 2 or 3 for the conversion.\n\nSo it's indeed surprising that you run out of memory. Is the dataset available somewhere ? or a subset maybe ?",
"This is a tokenized dataset I created for training a speech-language model with a few features (so it is not private but not easily available). I can send/upload a shard or two and you can copy them however many times you want so you can debug. this should give you something comparable to what I have, but will be easier than creating it yourself. so if you want that, let me know :)",
"Maybe you can measure the memory usage when loading 1 file with num_proc=1 ? This should already be helpful.\n\nMemory usage for tokenized data can be bigger than just text, for example the tokens type can be inferred as int64 and the lists offsets are int32",
"OK, I will try to do this in the near future. I am a little swamped at the moment. do you have a preferred tool?\n\nalso My data is just list of ints, there is no offsets",
"> so the json files are all loaded into memory before converting to arrow? or do they convert 1 json at a time and then they are realeased? I don't understand how 200GB worth of jsons fill a 378GB node's memory.\n\nHello! Is your query solved? I have the same confusion and would like to ask you for advice",
"no, the issue is still present. I converted the json files to parquet, but json seems to have a problem.\n\nUnfortunately i didn't have the time to try and profile the memory usage for 1 file. So if you want to do that, it will be great! ",
"My dataset is about image descriptions, stored as a 20MB JSON file on disk. However, I need to use the map function to preprocess the images, and after computation, the preprocessed dataset amounts to 70GB. My server has 122GB of RAM, but it still runs out of memory (OOM). This issue is very similar to yours.\n\nAfter some research during this period, I found that the map function does not perform disk mapping in memory while working. Using the command find /DataB/mjx -type f -mmin -10, I discovered that no temporary cache files were modified or created during program execution, meaning the data was continuously loaded into memory. After several attempts, I found that adding the parameter cache_file_name=\"your/path\" to the map function can enable memory-disk mapping. This is a strange setting, but after adding this parameter, the memory usage dropped to only 7GB, indicating that once the writer_batch_size worth of data is read into the disk cache, the corresponding data in memory is released. However, I don't think this is the intended behavior by the author, as memory-disk caching should have been enabled without needing this additional parameter.\n\nFinally, here is my map function call. I hope it helps you.\ntrain_data = train_data.map(process_fun, cache_file_name='./cache_file', remove_columns=['image_name', 'question_type', 'concern', 'question', 'candidate_answers', 'answer'])"
] |
2,986,612,934 |
Iterating over Image feature columns is extremely slow
|
open
|
We are trying to load datasets where the image column stores `PIL.PngImagePlugin.PngImageFile` images. However, iterating over these datasets is extremely slow.
What I have found:
1. It is the presence of the image column that causes the slowdown. Removing the column from the dataset results in blazingly fast (as expected) times
2. It is ~2x faster to iterate when the column contains a single image as opposed to a list of images i.e., the feature is a Sequence of Image objects. We often need multiple images per sample, so we need to work with a list of images
3. It is ~17x faster to store paths to PNG files and load them using `PIL.Image.open`, as opposed to iterating over a `Dataset` with an Image column, and ~30x faster compared to `Sequence` of `Image`s. See a simple script below with an openly available dataset.
It would be great to understand the standard practices for storing and loading multimodal datasets (image + text).
https://huggingface.co/docs/datasets/en/image_load seems a bit underdeveloped? (e.g., `dataset.decode` only works with `IterableDataset`, but it's not clear from the doc)
Thanks!
```python
from datasets import load_dataset, load_from_disk
from PIL import Image
from pathlib import Path
ds = load_dataset("getomni-ai/ocr-benchmark")
for idx, sample in enumerate(ds["test"]):
image = sample["image"]
image.save(f"/tmp/ds_files/images/image_{idx}.png")
ds.save_to_disk("/tmp/ds_columns")
# Remove the 'image' column
ds["test"] = ds["test"].remove_columns(["image"])
# Create image paths for each sample
image_paths = [f"images/image_{idx}.png" for idx in range(len(ds["test"]))]
# Add the 'image_path' column to the dataset
ds["test"] = ds["test"].add_column("image_path", image_paths)
# Save the updated dataset
ds.save_to_disk("/tmp/ds_files")
files_path = Path("/tmp/ds_files")
column_path = Path("/tmp/ds_columns")
# load and benchmark
ds_file = load_from_disk(files_path)
ds_column = load_from_disk(column_path)
import time
images_files = []
start = time.time()
for idx in range(len(ds_file["test"])):
image_path = files_path / ds_file["test"][idx]["image_path"]
image = Image.open(image_path)
images_files.append(image)
end = time.time()
print(f"Time taken to load images from files: {end - start} seconds")
# Time taken to load images from files: 1.2364635467529297 seconds
images_column = []
start = time.time()
for idx in range(len(ds_column["test"])):
images_column.append(ds_column["test"][idx]["image"])
end = time.time()
print(f"Time taken to load images from columns: {end - start} seconds")
# Time taken to load images from columns: 20.49347186088562 seconds
```
| 2025-04-10T19:00:54 | 2025-04-15T17:57:08 | null |
https://github.com/huggingface/datasets/issues/7508
| null | 7,508 | false |
[
"Hi ! Could it be because the `Image()` type in dataset does `image = Image.open(image_path)` and also `image.load()` which actually loads the image data in memory ? This is needed to avoid too many open files issues, see https://github.com/huggingface/datasets/issues/3985",
"Yes, that seems to be it. For my purposes, I've cast the column to `Image(decode=False)`, and only load the images when necessary, which is much much faster"
] |
2,984,309,806 |
Front-end statistical data quantity deviation
|
open
|
### Describe the bug
While browsing the dataset at https://huggingface.co/datasets/NeuML/wikipedia-20250123, I noticed that a dataset with nearly 7M entries was estimated to be only 4M in size—almost half the actual amount. According to the post-download loading and the dataset_info (https://huggingface.co/datasets/NeuML/wikipedia-20250123/blob/main/train/dataset_info.json), the true data volume is indeed close to 7M. This significant discrepancy could mislead users when sorting datasets by row count. Why not directly retrieve this information from dataset_info?
Not sure if this is the right place to report this bug, but leaving it here for the team's awareness.
| 2025-04-10T02:51:38 | 2025-04-15T12:54:51 | null |
https://github.com/huggingface/datasets/issues/7507
| null | 7,507 | false |
[
"Hi ! the format of this dataset is not supported by the Dataset Viewer. It looks like this dataset was saved using `save_to_disk()` which is meant for local storage / easy reload without compression, not for sharing online."
] |
2,981,687,450 |
HfHubHTTPError: 429 Client Error: Too Many Requests for URL when trying to access Fineweb-10BT on 4A100 GPUs using SLURM
|
open
|
### Describe the bug
I am trying to run some finetunings on 4 A100 GPUs using SLURM using axolotl training framework which in turn uses Huggingface's Trainer and Accelerate on [Fineweb-10BT](https://huggingface.co/datasets/HuggingFaceFW/fineweb), but I end up running into 429 Client Error: Too Many Requests for URL error when I call next(dataloader_iter). Funny is, that I can run some test fine tuning (for just 200 training steps) in 1 A100 GPU using SLURM. Is there any rate limiter set for querying dataset? I could run the fine tuning with the same settings (4 A100 GPUs in SLURM) last month.
### Steps to reproduce the bug
You would need a server installed with SLURM
1. Create conda environment
1.1 conda create -n example_env -c conda-forge gxx=11 python=3.10
1.2 conda activate example_env
1.3 pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu124
1.4 conda install nvidia/label/cuda-12.4.0::cuda-toolkit
1.5 Download flash_attn-2.7.4.post1+cu12torch2.5cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
1.6 pip3 install packaging
1.7 pip3 install ninja
1.8 pip3 install mlflow
1.9 Clone https://github.com/calvintanama/axolotl.git
1.10 `cd` to `axolotl`
1.11 pip3 install -e '.[deepspeed]'
2. Run the training
2.1. Create a folder called `config_run` in axolotl directory
2.2. Copy `config/phi3_pruned_extra_pretrain_22_29_bottleneck_residual_8_a100_4.yaml` to `config_run`
2.3. Change yaml file in the `config_run` accordingly
2.4. Change directory and conda environment name in `jobs/train_phi3_22_29_bottleneck_residual_8_a100_4_temp.sh`
2.5. `jobs/train_phi3_22_29_bottleneck_residual_8_a100_4_temp.sh`
### Expected behavior
This should not cause any error, but gotten
```
File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/accelerate/data_loader.py", line 552, in __iter__
[rank3]: current_batch = next(dataloader_iter)
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 701, in __next__
[rank3]: data = self._next_data()
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 757, in _next_data
[rank3]: data = self._dataset_fetcher.fetch(index) # may raise StopIteration
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py", line 33, in fetch
[rank3]: data.append(next(self.dataset_iter))
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/accelerate/data_loader.py", line 338, in __iter__
[rank3]: for element in self.dataset:
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 2266, in __iter__
[rank3]: for key, example in ex_iterable:
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 1866, in __iter__
[rank3]: for key, example in self.ex_iterable:
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 1084, in __iter__
[rank3]: yield from self._iter()
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 1263, in _iter
[rank3]: for key, transformed_example in outputs:
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 1258, in <genexpr>
[rank3]: outputs = (
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 1244, in iter_outputs
[rank3]: for i, key_example in inputs_iterator:
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 1106, in iter_batched_inputs
[rank3]: for key, example in iterator:
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 1866, in __iter__
[rank3]: for key, example in self.ex_iterable:
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 1535, in __iter__
[rank3]: for x in self.ex_iterable:
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/datasets/iterable_dataset.py", line 374, in __iter__
[rank3]: for key, pa_table in self.generate_tables_fn(**gen_kwags):
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/datasets/packaged_modules/parquet/parquet.py", line 90, in _generate_tables
[rank3]: if parquet_fragment.row_groups:
[rank3]: File "pyarrow/_dataset_parquet.pyx", line 386, in pyarrow._dataset_parquet.ParquetFileFragment.row_groups.__get__
[rank3]: File "pyarrow/_dataset_parquet.pyx", line 393, in pyarrow._dataset_parquet.ParquetFileFragment.metadata.__get__
[rank3]: File "pyarrow/_dataset_parquet.pyx", line 382, in pyarrow._dataset_parquet.ParquetFileFragment.ensure_complete_metadata
[rank3]: File "pyarrow/error.pxi", line 89, in pyarrow.lib.check_status
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/datasets/utils/file_utils.py", line 827, in read_with_retries
[rank3]: out = read(*args, **kwargs)
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/huggingface_hub/hf_file_system.py", line 1013, in read
[rank3]: return super().read(length)
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/fsspec/spec.py", line 1941, in read
[rank3]: out = self.cache._fetch(self.loc, self.loc + length)
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/fsspec/caching.py", line 234, in _fetch
[rank3]: self.cache = self.fetcher(start, end) # new block replaces old
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/huggingface_hub/hf_file_system.py", line 976, in _fetch_range
[rank3]: hf_raise_for_status(r)
[rank3]: File "/home/hk-project-test-p0023745/cd7437/miniconda3/envs/llmpruning_train_temp/lib/python3.10/site-packages/huggingface_hub/utils/_http.py", line 482, in hf_raise_for_status
[rank3]: raise _format(HfHubHTTPError, str(e), response) from e
[rank3]: huggingface_hub.errors.HfHubHTTPError: 429 Client Error: Too Many Requests for url: https://huggingface.co/datasets/HuggingFaceFW/fineweb/resolve/0f039043b23fe1d4eed300b504aa4b4a68f1c7ba/sample/10BT/006_00000.parquet
```
### Environment info
- datasets 3.5.0
- torch 2.5.1
- transformers 4.46.2
| 2025-04-09T06:32:04 | 2025-06-29T06:04:59 | null |
https://github.com/huggingface/datasets/issues/7506
| null | 7,506 | false |
[
"Hi ! make sure to be logged in with your HF account (e.g. using `huggingface-cli login` or passing `token=` to `load_dataset()`), otherwise you'll get rate limited at one point",
"Hey @calvintanama! Just building on what @lhoestq mentioned above — I ran into similar issues in multi-GPU SLURM setups and here’s what worked for me...\n\nThis 429 Client Error: Too Many Requests comes from the Hugging Face Hub’s rate limiting, which restricts unauthenticated or high-volume access (especially in multi-GPU/distributed setups like SLURM).\n\nAs @lhoestq mentioned, the solution is to make sure you are authenticated with the Hugging Face Hub in every process (especially on each GPU/worker node). You can do this by:\n\nRunning huggingface-cli login (interactive)\n\nOr passing your token explicitly:\n\n```python\nload_dataset(\"HuggingFaceFW/fineweb\", token=\"hf_your_token_here\")\n# If you’re using a SLURM cluster, ensure every node/process receives access to the token via env var:\n```\n\n```bash\nexport HF_TOKEN=hf_your_token_here\n```\n\nand then in Python:\n```python\nfrom datasets import load_dataset\nload_dataset(\"HuggingFaceFW/fineweb\", token=os.environ[\"HF_TOKEN\"])\n```\nAlso consider downloading the dataset beforehand with load_dataset(..., streaming=False) and storing it locally if you're repeatedly training with it."
] |
2,979,926,156 |
HfHubHTTPError: 403 Forbidden: None. Cannot access content at: https://hf.co/api/s3proxy
|
open
|
I have already logged in Huggingface using CLI with my valid token. Now trying to download the datasets using following code:
from transformers import WhisperProcessor, WhisperForConditionalGeneration, WhisperTokenizer, Trainer, TrainingArguments, DataCollatorForSeq2Seq
from datasets import load_dataset, DatasetDict, Audio
def load_and_preprocess_dataset():
dataset = load_dataset("mozilla-foundation/common_voice_17_0", "bn")
dataset = dataset.remove_columns(["accent", "age", "client_id", "down_votes", "gender", "locale", "segment", "up_votes"])
dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
dataset = dataset["train"].train_test_split(test_size=0.1)
dataset = DatasetDict({
"train": dataset["train"],
"test": dataset["test"]
})
return dataset
load_and_preprocess_dataset()
I am getting following error:
Downloading data: 100%
25/25 [00:01<00:00, 25.31files/s]
---------------------------------------------------------------------------
HTTPError Traceback (most recent call last)
File ~/github/bangla-asr/.venv/lib/python3.11/site-packages/huggingface_hub/utils/_http.py:409, in hf_raise_for_status(response, endpoint_name)
408 try:
--> 409 response.raise_for_status()
410 except HTTPError as e:
File ~/github/bangla-asr/.venv/lib/python3.11/site-packages/requests/models.py:1024, in Response.raise_for_status(self)
1023 if http_error_msg:
-> 1024 raise HTTPError(http_error_msg, response=self)
HTTPError: 403 Client Error: BlockSIEL for url: https://hf.co/api/s3proxy?GET=https%3A%2F%2Fhf-hub-lfs-us-east-1.s3.us-east-1.amazonaws.com%2Frepos%2Fa3%2F86%2Fa386bf65687d8a6928c1ea57c383aa3faade32f5171150e25af3fc1cfc273db8%2F67f1ac9cabd539bfbff3acbc549b60647833a250dc638866f22bf1b64e68806d%3FX-Amz-Algorithm%3DAWS4-HMAC-SHA256%26X-Amz-Content-Sha256%3DUNSIGNED-PAYLOAD%26X-Amz-Credential%3DAKIA2JU7TKAQLC2QXPN7%252F20250408%252Fus-east-1%252Fs3%252Faws4_request%26X-Amz-Date%3D20250408T134345Z%26X-Amz-Expires%3D3600%26X-Amz-Signature%3D621e731d4fd6d08afbf568379797746ab8e2b853b6728ff5e1122fef6e56880b%26X-Amz-SignedHeaders%3Dhost%26response-content-disposition%3Dinline%253B%2520filename%252A%253DUTF-8%2527%2527bn_validated_1.tar%253B%2520filename%253D%2522bn_validated_1.tar%2522%253B%26response-content-type%3Dapplication%252Fx-tar%26x-id%3DGetObject&HEAD=https%3A%2F%2Fhf-hub-lfs-us-east-1.s3.us-east-1.amazonaws.com%2Frepos%2Fa3%2F86%2Fa386bf65687d8a6928c1ea57c383aa3faade32f5171150e25af3fc1cfc273db8%2F67f1ac9cabd539bfbff3acbc549b60647833a250dc638866f22bf1b64e68806d%3FX-Amz-Algorithm%3DAWS4-HMAC-SHA256%26X-Amz-Content-Sha256%3DUNSIGNED-PAYLOAD%26X-Amz-Credential%3DAKIA2JU7TKAQLC2QXPN7%252F20250408%252Fus-east-1%252Fs3%252Faws4_request%26X-Amz-Date%3D20250408T134345Z%26X-Amz-Expires%3D3600%26X-Amz-Signature%3D15254fb79d30b0dc36b94a28138e675e0e00bb475b8a3ae774418500b095a661%26X-Amz-SignedHeaders%3Dhost&sign=eyJhbGciOiJIUzI1NiJ9.eyJyZWRpcmVjdF9kb21haW4iOiJoZi1odWItbGZzLXVzLWVhc3QtMS5zMy51cy1lYXN0LTEuYW1hem9uYXdzLmNvbSIsImlhdCI6MTc0NDExOTgyNSwiZXhwIjoxNzQ0MjA2MjI1LCJpc3MiOiJodHRwczovL2h1Z2dpbmdmYWNlLmNvIn0.5sJzudFDU3SmOdOLlwmQCOfQFf2r7y9590HoX8WBkRk
The above exception was the direct cause of the following exception:
HfHubHTTPError Traceback (most recent call last)
Cell In[16], line 15
9 dataset = DatasetDict({
10 "train": dataset["train"],
11 "test": dataset["test"]
12 })
13 return dataset
---> 15 load_and_preprocess_dataset()
17 # def setup_model():
18 # processor = WhisperProcessor.from_pretrained("openai/whisper-base")
...
475 range_header = response.request.headers.get("Range")
HfHubHTTPError: 403 Forbidden: None.
Cannot access content at: https://hf.co/api/s3proxy?GET=https%3A%2F%2Fhf-hub-lfs-us-east-1.s3.us-east-1.amazonaws.com%2Frepos%2Fa3%2F86%2Fa386bf6568724a6928c1ea57c383aa3faade32f5171150e25af3fc1cfc273db8%2F67f1ac9cabd539bfbff3acbc549b60647833a250dc638786f22bf1b64e68806d%3FX-Amz-Algorithm%3DAWS4-HMAC-SHA256%26X-Amz-Content-Sha256%3DUNSIGNED-PAYLOAD%26X-Amz-Credential%3DAKIA2JU7TKAQLC2QXPN7%252F20250408%252Fus-east-1%252Fs3%252Faws4_request%26X-Amz-Date%3D20250408T134345Z%26X-Amz-Expires%3D3600%26X-Amz-Signature%3D621e731d4fd6d08afbf568379797746ab394b853b6728ff5e1122fef6e56880b%26X-Amz-SignedHeaders%3Dhost%26response-content-disposition%3Dinline%253B%2520filename%252A%253DUTF-8%2527%2527bn_validated_1.tar%253B%2520filename%253D%2522bn_validated_1.tar%2522%253B%26response-content-type%3Dapplication%252Fx-tar%26x-id%3DGetObject&HEAD=https%3A%2F%2Fhf-hub-lfs-us-east-1.s3.us-east-1.amazonaws.com%2Frepos%2Fa3%2F86%2Fa386bf65687ab76928c1ea57c383aa3faade32f5171150e25af3fc1cfc273db8%2F67f1ac9cabd539bfbff3acbc549b60647833a250d2338866f222f1b64e68806d%3FX-Amz-Algorithm%3DAWS4-HMAC-SHA256%26X-Amz-Content-Sha256%3DUNSIGNED-PAYLOAD%26X-Amz-Credential%3DAKIA2JU7TKAQLC2QXPN7%252F20250408%252Fus-east-1%252Fs3%252Faws4_request%26X-Amz-Date%3D20250408T134345Z%26X-Amz-Expires%3D3600%26X-Amz-Signature%3D15254fb79d30b0dc36b94a28138e675e0e00bb475b8a3ae774418500b095a661%26X-Amz-SignedHeaders%3Dhost&sign=eyJhbGciOiJIUzI1NiJ9.eyJyZWRpcmVjds9kb21haW4iOiJoZi1odWItbGZzLXVzLWVhc3QtMS5zMy51cy1lYXN0LTEuYW1hem9uYXdzLmNvbSIsImlhdCI6MTc0NDExOT2yNSwiZXhwIjoxNzQ0MjA2MjI1LCJpc3MiOiJodHRwczovL2h1Z2dpbmdmYWNlLmNvIn0.5sJzudFDU3SmOdOLlwmQdOfQFf2r7y9590HoX8WBkRk.
Make sure your token has the correct permissions.
**What's wrong with the code?** Please note that the error is happening only when I am running from my office network due to probably proxy. Which URL, I need to take a proxy exception?
| 2025-04-08T14:08:40 | 2025-04-08T14:08:40 | null |
https://github.com/huggingface/datasets/issues/7505
| null | 7,505 | false |
[] |
2,979,410,641 |
BuilderConfig ParquetConfig(...) doesn't have a 'use_auth_token' key.
|
open
|
### Describe the bug
Trying to run the following fine-tuning script (based on this page [here](https://github.com/huggingface/instruction-tuned-sd)):
```
! accelerate launch /content/instruction-tuned-sd/finetune_instruct_pix2pix.py \
--pretrained_model_name_or_path=${MODEL_ID} \
--dataset_name=${DATASET_NAME} \
--use_ema \
--enable_xformers_memory_efficient_attention \
--resolution=512 --random_flip \
--train_batch_size=2 --gradient_accumulation_steps=4 --gradient_checkpointing \
--max_train_steps=500 \
--checkpointing_steps=25 --checkpoints_total_limit=1 \
--learning_rate=5e-05 --max_grad_norm=1 --lr_warmup_steps=20 \
--conditioning_dropout_prob=0.1 \
--mixed_precision=fp16 \
--seed=42 \
--output_dir=${OUTPUT_DIR} \
--original_image_column=before \
--edit_prompt=prompt \
--edited_image=after
```
but I keep getting the following error:
```
Traceback (most recent call last):
File "/content/instruction-tuned-sd/finetune_instruct_pix2pix.py", line 1137, in <module>
main()
File "/content/instruction-tuned-sd/finetune_instruct_pix2pix.py", line 652, in main
dataset = load_dataset(
^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/dist-packages/datasets/load.py", line 2129, in load_dataset
builder_instance = load_dataset_builder(
^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/dist-packages/datasets/load.py", line 1886, in load_dataset_builder
builder_instance: DatasetBuilder = builder_cls(
^^^^^^^^^^^^
File "/usr/local/lib/python3.11/dist-packages/datasets/builder.py", line 342, in __init__
self.config, self.config_id = self._create_builder_config(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/dist-packages/datasets/builder.py", line 590, in _create_builder_config
raise ValueError(f"BuilderConfig {builder_config} doesn't have a '{key}' key.")
ValueError: BuilderConfig ParquetConfig(name='default', version=0.0.0, data_dir=None, data_files={'train': ['data/train-*']}, description=None, batch_size=None, columns=None, features=None, filters=None) doesn't have a 'use_auth_token' key.
Traceback (most recent call last):
File "/usr/local/bin/accelerate", line 10, in <module>
sys.exit(main())
^^^^^^
```
Any ideas? `datasets` version should be `3.2.0`.
### Steps to reproduce the bug
Just running the script above.
### Expected behavior
No errors
### Environment info
Python 3.11.11
datasets==3.2.0
| 2025-04-08T10:55:03 | 2025-06-28T09:18:09 | null |
https://github.com/huggingface/datasets/issues/7504
| null | 7,504 | false |
[
"I encountered the same error, have you resolved it?",
"Hi ! `use_auth_token` has been deprecated and removed some time ago. You should use `token` instead in `load_dataset()`",
"Hi @lhoestq, I'd like to take this up.\n\nAs discussed in #7504, the issue arises when `use_auth_token` is passed to `load_dataset`, which forwards it to the config's `__init__`, where it's no longer a valid key.\n\nTo address this, I’ll intercept and strip `use_auth_token` inside `load_dataset()` (similar to how we handle `trust_remote_code`). A warning will be logged, and users will be encouraged to use `token` instead.\n\nThis avoids breaking older scripts that still use `use_auth_token`."
] |
2,978,512,625 |
Inconsistency between load_dataset and load_from_disk functionality
|
open
|
## Issue Description
I've encountered confusion when using `load_dataset` and `load_from_disk` in the datasets library. Specifically, when working offline with the gsm8k dataset, I can load it using a local path:
```python
import datasets
ds = datasets.load_dataset('/root/xxx/datasets/gsm8k', 'main')
```
output:
```text
DatasetDict({
train: Dataset({
features: ['question', 'answer'],
num_rows: 7473
})
test: Dataset({
features: ['question', 'answer'],
num_rows: 1319
})
})
```
This works as expected. However, after processing the dataset (converting answer format from #### to \boxed{})
```python
import datasets
ds = datasets.load_dataset('/root/xxx/datasets/gsm8k', 'main')
ds_train = ds['train']
ds_test = ds['test']
import re
def convert(sample):
solution = sample['answer']
solution = re.sub(r'####\s*(\S+)', r'\\boxed{\1}', solution)
sample = {
'problem': sample['question'],
'solution': solution
}
return sample
ds_train = ds_train.map(convert, remove_columns=['question', 'answer'])
ds_test = ds_test.map(convert,remove_columns=['question', 'answer'])
```
I saved it using save_to_disk:
```python
from datasets.dataset_dict import DatasetDict
data_dict = DatasetDict({
'train': ds_train,
'test': ds_test
})
data_dict.save_to_disk('/root/xxx/datasets/gsm8k-new')
```
But now I can only load it using load_from_disk:
```python
new_ds = load_from_disk('/root/xxx/datasets/gsm8k-new')
```
output:
```text
DatasetDict({
train: Dataset({
features: ['problem', 'solution'],
num_rows: 7473
})
test: Dataset({
features: ['problem', 'solution'],
num_rows: 1319
})
})
```
Attempting to use load_dataset produces unexpected results:
```python
new_ds = load_dataset('/root/xxx/datasets/gsm8k-new')
```
output:
```text
DatasetDict({
train: Dataset({
features: ['_data_files', '_fingerprint', '_format_columns', '_format_kwargs', '_format_type', '_output_all_columns', '_split'],
num_rows: 1
})
test: Dataset({
features: ['_data_files', '_fingerprint', '_format_columns', '_format_kwargs', '_format_type', '_output_all_columns', '_split'],
num_rows: 1
})
})
```
Questions
1. Why is it designed such that after using `save_to_disk`, the dataset cannot be loaded with `load_dataset`? For small projects with limited code, it might be relatively easy to change all instances of `load_dataset` to `load_from_disk`. However, for complex frameworks like TRL or lighteval, diving into the framework code to change `load_dataset` to `load_from_disk` is extremely tedious and error-prone.
Additionally, `load_from_disk` cannot load datasets directly downloaded from the hub, which means that if you need to modify a dataset, you have to choose between using `load_from_disk` or `load_dataset`. This creates an unnecessary dichotomy in the API and complicates workflow when working with modified datasets.
2. What's the recommended approach for this use case? Should I manually process my gsm8k-new dataset to make it compatible with load_dataset? Is there a standard way to convert between these formats?
thanks~
| 2025-04-08T03:46:22 | 2025-06-28T08:51:16 | null |
https://github.com/huggingface/datasets/issues/7503
| null | 7,503 | false |
[
"Hi ! you can find more info here: https://github.com/huggingface/datasets/issues/5044#issuecomment-1263714347\n\n> What's the recommended approach for this use case? Should I manually process my gsm8k-new dataset to make it compatible with load_dataset? Is there a standard way to convert between these formats?\n\nYou can use push_to_hub() or to_parquet() for example",
"Hi @zzzzzec & @lhoestq 👋\n\nThanks for raising and discussing this — I've submitted a patch that improves this exact scenario."
] |
2,977,453,814 |
`load_dataset` of size 40GB creates a cache of >720GB
|
closed
|
Hi there,
I am trying to load a dataset from the Hugging Face Hub and split it into train and validation splits. Somehow, when I try to do it with `load_dataset`, it exhausts my disk quota. So, I tried manually downloading the parquet files from the hub and loading them as follows:
```python
ds = DatasetDict(
{
"train": load_dataset(
"parquet",
data_dir=f"{local_dir}/{tok}",
cache_dir=cache_dir,
num_proc=min(12, os.cpu_count()), # type: ignore
split=ReadInstruction("train", from_=0, to=NUM_TRAIN, unit="abs"), # type: ignore
),
"validation": load_dataset(
"parquet",
data_dir=f"{local_dir}/{tok}",
cache_dir=cache_dir,
num_proc=min(12, os.cpu_count()), # type: ignore
split=ReadInstruction("train", from_=NUM_TRAIN, unit="abs"), # type: ignore
)
}
)
```
which still strangely creates 720GB of cache. In addition, if I remove the raw parquet file folder (`f"{local_dir}/{tok}"` in this example), I am not able to load anything. So, I am left wondering what this cache is doing. Am I missing something? Is there a solution to this problem?
Thanks a lot in advance for your help!
A related issue: https://github.com/huggingface/transformers/issues/10204#issue-809007443.
---
Python: 3.11.11
datasets: 3.5.0
| 2025-04-07T16:52:34 | 2025-04-15T15:22:12 | 2025-04-15T15:22:11 |
https://github.com/huggingface/datasets/issues/7502
| null | 7,502 | false |
[
"Hi ! Parquet is a compressed format. When you load a dataset, it uncompresses the Parquet data into Arrow data on your disk. That's why you can indeed end up with 720GB of uncompressed data on disk. The uncompression is needed to enable performant dataset objects (especially for random access).\n\nTo save some storage you can instead load the dataset with `streaming=True`. This way you get an `IterableDataset` that reads the Parquet data iteratively without ever writing to disk.\n\nPS: `ReadInstruction` might not be implemented for `streaming=True`, if it's the case you can use `ds.take()` and `ds.skip()` instead",
"Hi @lhoestq, thanks a lot for your answer. This makes perfect sense. I will try using the streaming mode. Closing the issue."
] |
2,976,721,014 |
Nested Feature raises ArrowNotImplementedError: Unsupported cast using function cast_struct
|
closed
|
### Describe the bug
`datasets.Features` seems to be unable to handle json file that contains fields of `list[dict]`.
### Steps to reproduce the bug
```json
// test.json
{"a": 1, "b": [{"c": 2, "d": 3}, {"c": 4, "d": 5}]}
{"a": 5, "b": [{"c": 7, "d": 8}, {"c": 9, "d": 10}]}
```
```python
import json
from datasets import Dataset, Features, Value, Sequence, load_dataset
annotation_feature = Features({
"a": Value("int32"),
"b": Sequence({
"c": Value("int32"),
"d": Value("int32"),
}),
})
annotation_dataset = load_dataset(
"json",
data_files="test.json",
features=annotation_feature
)
```
```
ArrowNotImplementedError: Unsupported cast from list<item: struct<c: int32, d: int32>> to struct using function cast_struct
The above exception was the direct cause of the following exception:
DatasetGenerationError Traceback (most recent call last)
Cell In[46], line 11
2 from datasets import Dataset, Features, Value, Sequence, load_dataset
4 annotation_feature = Features({
5 "a": Value("int32"),
6 "b": Sequence({
(...) 9 }),
10 })
---> 11 annotation_dataset = load_dataset(
12 "json",
13 data_files="test.json",
14 features=annotation_feature
15 )
```
### Expected behavior
A `datasets.Datasets` instance should be initialized.
### Environment info
- `datasets` version: 3.5.0
- Platform: Linux-6.11.0-21-generic-x86_64-with-glibc2.39
- Python version: 3.11.11
- `huggingface_hub` version: 0.30.1
- PyArrow version: 19.0.1
- Pandas version: 2.2.3
- `fsspec` version: 2024.12.0
| 2025-04-07T12:35:39 | 2025-04-07T12:43:04 | 2025-04-07T12:43:03 |
https://github.com/huggingface/datasets/issues/7501
| null | 7,501 | false |
[
"Solved by the default `load_dataset(features)` parameters. Do not use `Sequence` for the `list` in `list[any]` json schema, just simply use `[]`. For example, `\"b\": Sequence({...})` fails but `\"b\": [{...}]` works fine."
] |
2,974,841,921 |
Make `with_format` correctly indicate that a `Dataset` is compatible with PyTorch's `Dataset` class
|
open
|
### Feature request
Currently `datasets` does not correctly indicate to the Python type-checker (e.g. `pyright` / `Pylance`) that the output of `with_format` is compatible with PyTorch's `Dataloader` since it does not indicate that the HuggingFace `Dataset` is compatible with the PyTorch `Dataset` class. It would be great if we could get the typing to work nicely.
### Motivation
To avoid casting types in our Python code.
### Your contribution
I would be happy to contribute a PR if this is something that may be accepted and could work with the current approach.
This doesn't have to be for just PyTorch, but I imagine that the same thing would be useful for `tensorflow` and such, but we only have a need for PyTorch at this stage.
| 2025-04-06T09:56:09 | 2025-04-15T12:57:39 | null |
https://github.com/huggingface/datasets/issues/7500
| null | 7,500 | false |
[
"Does the torch `DataLoader` really require the dataset to be a subclass of `torch.utils.data.Dataset` ? Or is there a simpler type we could use ?\n\nPS: also note that a dataset without `with_format()` can also be used in a torch `DataLoader` . Calling `with_format(\"torch\")` simply makes the output of the dataset torch Tensors in an efficient way."
] |
2,973,489,126 |
Added cache dirs to load and file_utils
|
closed
|
When adding "cache_dir" to datasets.load_dataset, the cache_dir gets lost in the function calls, changing the cache dir to the default path. This fixes a few of these instances.
| 2025-04-04T22:36:04 | 2025-05-07T14:07:34 | 2025-05-07T14:07:34 |
https://github.com/huggingface/datasets/pull/7499
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7499",
"html_url": "https://github.com/huggingface/datasets/pull/7499",
"diff_url": "https://github.com/huggingface/datasets/pull/7499.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7499.patch",
"merged_at": null
}
| 7,499 | true |
[
"hi ! the `hf_hub_download` cache_dir is a different cache directory than the one for `datasets`.\r\n\r\n`hf_hub_download` uses the `huggingface_hub` cache which is located in by default in `~/.cache/huggingface/hub`, while `datasets` uses a different cache for Arrow files and map() results `~/.cache/huggingface/datasets`",
"Is there a way to change the default cache directory for both of these on calling load_dataset? Currently, cache_dir makes dealing with where I want files to go a bit confusing as the documentation doesn't mention it only relocates.../datasets and not .../hub.",
"You can set `HF_HOME` which is the common parent directory for those two caches. Or individually `HF_DATASETS_CACHE` and `HF_HUB_CACHE`",
"Got it. Can this be added to the documentation for load_dataset and related functions to avoid confusion with cache_dir?",
"done in https://github.com/huggingface/datasets/pull/7532 :)"
] |
2,969,218,273 |
Extreme memory bandwidth.
|
open
|
### Describe the bug
When I use hf datasets on 4 GPU with 40 workers I get some extreme memory bandwidth of constant ~3GB/s.
However, if I wrap the dataset in `IterableDataset`, this issue is gone and the data also loads way faster (4x faster training on 1 worker).
It seems like the workers don't share memory and basically duplicate the data 4x40.
### Steps to reproduce the bug
Trainer arguments:
```
dataloader_pin_memory=True,
dataloader_num_workers=40,
dataloader_prefetch_factor=2,
dataloader_persistent_workers=True,
```
Call trainer:
```
trainer = Trainer(
model=model,
args=train_args,
train_dataset=load_from_disk('..').with_fromat('torch'),
)
```
The dataset has 600GB and consists of 1225 files.
### Expected behavior
The optimal bandwidth should be 100MB/s to keep up with GPU.
### Environment info
Linux
Python 3.11
datasets==3.2.0
| 2025-04-03T11:09:08 | 2025-04-03T11:11:22 | null |
https://github.com/huggingface/datasets/issues/7498
| null | 7,498 | false |
[] |
2,968,553,693 |
How to convert videos to images?
|
open
|
### Feature request
Does someone know how to return the images from videos?
### Motivation
I am trying to use openpi(https://github.com/Physical-Intelligence/openpi) to finetune my Lerobot dataset(V2.0 and V2.1). I find that although the codedaset is v2.0, they are different. It seems like Lerobot V2.0 has two version, one is data include images infos and another one is separate to data and videos.
Does someone know how to return the images from videos?
| 2025-04-03T07:08:39 | 2025-04-15T12:35:15 | null |
https://github.com/huggingface/datasets/issues/7497
| null | 7,497 | false |
[
"Hi ! there is some documentation here on how to read video frames: https://huggingface.co/docs/datasets/video_load"
] |
2,967,345,522 |
Json builder: Allow features to override problematic Arrow types
|
open
|
### Feature request
In the JSON builder, use explicitly requested feature types before or while converting to Arrow.
### Motivation
Working with JSON datasets is really hard because of Arrow. At the very least, it seems like it should be possible to work-around these problems by explicitly setting problematic columns's types. But it seems like this is not possible because the features are only used *after* converting to arrow.
Here's a simple example where the Arrow error could potentially be avoided by converting the column to a string: https://colab.research.google.com/drive/16QHRdbUwKSrpwVfGwu8V8AHr8v2dv0dt?usp=sharing
### Your contribution
Maybe with some guidance. I'm not very familiar with arrow or pandas.
| 2025-04-02T19:27:16 | 2025-04-15T13:06:09 | null |
https://github.com/huggingface/datasets/issues/7496
| null | 7,496 | false |
[
"Hi ! It would be cool indeed, currently the JSON data are generally loaded here: \n\nhttps://github.com/huggingface/datasets/blob/90e5bf8a8599b625d6103ee5ac83b98269991141/src/datasets/packaged_modules/json/json.py#L137-L140\n\nMaybe we can pass a Arrow `schema` to avoid errors ?"
] |
2,967,034,060 |
Columns in the dataset obtained though load_dataset do not correspond to the one in the dataset viewer since 3.4.0
|
closed
|
### Describe the bug
I have noticed that on my dataset named [BrunoHays/Accueil_UBS](https://huggingface.co/datasets/BrunoHays/Accueil_UBS), since the version 3.4.0, every column except audio is missing when I load the dataset.
Interestingly, the dataset viewer still shows the correct columns
### Steps to reproduce the bug
```python
from datasets import load_dataset
ds = load_dataset("BrunoHays/Accueil_UBS", streaming=True)
print(next(iter(ds["test"])).keys())
```
With datasets >= 3.4.0:
-> dict_keys(['audio'])
With datasets == 3.3.2:
-> dict_keys(['audio', 'id', 'speaker', 'sentence', 'raw_sentence', 'start_timestamp', 'end_timestamp', 'overlap'])
### Expected behavior
All the columns should be present
### Environment info
- `datasets` version: 3.3.2
- Platform: macOS-14.6.1-x86_64-i386-64bit
- Python version: 3.10.15
- `huggingface_hub` version: 0.30.1
- PyArrow version: 16.1.0
- Pandas version: 1.5.3
- `fsspec` version: 2023.10.0
| 2025-04-02T17:01:11 | 2025-07-02T23:24:57 | 2025-07-02T23:24:57 |
https://github.com/huggingface/datasets/issues/7495
| null | 7,495 | false |
[
"Hi, the dataset viewer shows all the possible columns and their types, but `load_dataset()` iterates through all the columns that you defined. It seems that you only have one column (‘audio’) defined in your dataset because when I ran `print(ds.column_names)`, the only name I got was “audio”. You need to clearly define all the other features of the dataset as columns to enable your original code to work. Furthermore, you can run this code to print out all the features of your dataset: \n```python\nfrom datasets import load_dataset_builder\nds_builder = load_dataset_builder(\"BrunoHays/Accueil_UBS\")\nprint(ds_builder.info.features)\n```\n",
"@phoebecd \nGood catch, even in datasets<3.4.0, the only feature is \"audio\".\nThis datasets follows the [audio folder](https://huggingface.co/docs/datasets/en/audio_dataset#audiofolder) structure with metadata.csv.\nMaybe I missed something or there is a bug when having and audio_folder with a metadata file\n\nWhat do you think @lhoestq ?",
"I opened a PR to fix the issue :) https://huggingface.co/datasets/BrunoHays/Accueil_UBS/discussions/2\n\nWe expect the metadata file to be in the <split>/ folder now to allow one CSV metadata file per split. But in the PR I just added a manual configuration instead of moving the file and updating all the relative paths it contains."
] |
2,965,347,685 |
Broken links in pdf loading documentation
|
closed
|
### Describe the bug
Hi, just a couple of small issues I ran into while reading the docs for [loading pdf data](https://huggingface.co/docs/datasets/main/en/document_load):
1. The link for the [`Create a pdf dataset`](https://huggingface.co/docs/datasets/main/en/document_load#pdffolder) points to https://huggingface.co/docs/datasets/main/en/pdf_dataset instead of https://huggingface.co/docs/datasets/main/en/document_dataset and hence gives a 404 error.
2. At the top of the page, it's mentioned that to work with pdf datasets we need to have the `pdfplumber` package installed but the link to its installation guide points to `pytorch/vision` [installation instructions](https://github.com/pytorch/vision#installation) instead of `pdfplumber`'s [guide](https://github.com/jsvine/pdfplumber#installation)
I love the work on enabling pdf dataset support and these small tweaks would help everyone navigate the docs better. Thanks!
### Steps to reproduce the bug
The issue is on the [Load Document Data](https://huggingface.co/docs/datasets/main/en/document_load) page of the datasets docs.
### Expected behavior
1. For solving the first issue, I went through the [source .mdx code](https://github.com/huggingface/datasets/blob/main/docs/source/document_load.mdx?plain=1#L188) of the datasets docs and found that the link is pointing to `./pdf_dataset` instead of `./document_dataset`
2. For the second issue, I went through the [source .mdx code](https://github.com/huggingface/datasets/blob/main/docs/source/document_load.mdx?plain=1#L13) of the datasets docs and found that the link is `pytorch/vision` [installation instructions](https://github.com/pytorch/vision#installation) instead of `pdfplumber`'s [guide](https://github.com/jsvine/pdfplumber#installation)
Just replacing these two links should fix the bugs
### Environment info
datasets v3.5.0 (main at the time of writing)
| 2025-04-02T06:45:22 | 2025-04-15T13:36:25 | 2025-04-15T13:36:04 |
https://github.com/huggingface/datasets/issues/7494
| null | 7,494 | false |
[
"thanks for reporting ! I fixed the links, the docs will be updated in the next release"
] |
2,964,025,179 |
push_to_hub does not upload videos
|
open
|
### Describe the bug
Hello,
I would like to upload a video dataset (some .mp4 files and some segments within them), i.e. rows correspond to subsequences from videos. Videos might be referenced by several rows.
I created a dataset locally and it references the videos and the video readers can read them correctly. I use push_to_hub() to upload the dataset to the hub.
Expectation: A user uses `load_dataset` and can load the videos.
However, the videos seem to be just referenced via paths on the computer and not uploaded to the hub. Therefore a target user cannot load the videos in the dataset.
### Steps to reproduce the bug
1. create a video dataset with paths e.g. { ["videos"]: [path1, path2, ...]}
2. dataset.push_to_hub
3. on a different computer (or same pc if relative paths are used in a different folder):
```
dataset = load_dataset("siplab/egosim", split="train")
video = dataset[0]["video_head"]
```
3. will fail
### Expected behavior
Expectation: A user uses `load_dataset` and can load the videos.
### Environment info
datasets 3.1.0
Python 3.8.18
| 2025-04-01T17:00:20 | 2025-04-15T12:34:23 | null |
https://github.com/huggingface/datasets/issues/7493
| null | 7,493 | false |
[
"Hi ! the `Video` type is still experimental, and in particular `push_to_hub` doesn't upload videos at the moment (only the paths).\n\nThere is an open question to either upload the videos inside the Parquet files, or rather have them as separate files (which is great to enable remote seeking/streaming)"
] |
2,959,088,568 |
Closes #7457
|
closed
|
This PR updates the documentation to include the HF_DATASETS_CACHE environment variable, which allows users to customize the cache location for datasets—similar to HF_HUB_CACHE for models.
| 2025-03-30T20:41:20 | 2025-04-13T22:05:07 | 2025-04-13T22:05:07 |
https://github.com/huggingface/datasets/pull/7492
|
{
"url": "https://api.github.com/repos/huggingface/datasets/pulls/7492",
"html_url": "https://github.com/huggingface/datasets/pull/7492",
"diff_url": "https://github.com/huggingface/datasets/pull/7492.diff",
"patch_url": "https://github.com/huggingface/datasets/pull/7492.patch",
"merged_at": null
}
| 7,492 | true |
[
"This PR fixes issue #7457"
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.