ayushexel/colbert-MiniLM-L6-H384-uncased-1-neg-1-epoch-gooaq-1995000

PyLate model based on nreimers/MiniLM-L6-H384-uncased

This is a PyLate model finetuned from nreimers/MiniLM-L6-H384-uncased. It maps sentences & paragraphs to sequences of 128-dimensional dense vectors and can be used for semantic textual similarity using the MaxSim operator.

Model Details

Model Description

• Model Type: PyLate model
• Base model: nreimers/MiniLM-L6-H384-uncased
• Document Length: 180 tokens
• Query Length: 32 tokens
• Output Dimensionality: 128 tokens
• Similarity Function: MaxSim

Model Sources

• Documentation: PyLate Documentation
• Repository: PyLate on GitHub
• Hugging Face: PyLate models on Hugging Face

Full Model Architecture

ColBERT(
  (0): Transformer({'max_seq_length': 31, 'do_lower_case': False}) with Transformer model: BertModel 
  (1): Dense({'in_features': 384, 'out_features': 128, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity'})
)

Usage

First install the PyLate library:

pip install -U pylate

Retrieval

PyLate provides a streamlined interface to index and retrieve documents using ColBERT models. The index leverages the Voyager HNSW index to efficiently handle document embeddings and enable fast retrieval.

Indexing documents

First, load the ColBERT model and initialize the Voyager index, then encode and index your documents:

from pylate import indexes, models, retrieve

# Step 1: Load the ColBERT model
model = models.ColBERT(
    model_name_or_path=ayushexel/colbert-MiniLM-L6-H384-uncased-1-neg-1-epoch-gooaq-1995000,
)

# Step 2: Initialize the Voyager index
index = indexes.Voyager(
    index_folder="pylate-index",
    index_name="index",
    override=True,  # This overwrites the existing index if any
)

# Step 3: Encode the documents
documents_ids = ["1", "2", "3"]
documents = ["document 1 text", "document 2 text", "document 3 text"]

documents_embeddings = model.encode(
    documents,
    batch_size=32,
    is_query=False,  # Ensure that it is set to False to indicate that these are documents, not queries
    show_progress_bar=True,
)

# Step 4: Add document embeddings to the index by providing embeddings and corresponding ids
index.add_documents(
    documents_ids=documents_ids,
    documents_embeddings=documents_embeddings,
)

Note that you do not have to recreate the index and encode the documents every time. Once you have created an index and added the documents, you can re-use the index later by loading it:

# To load an index, simply instantiate it with the correct folder/name and without overriding it
index = indexes.Voyager(
    index_folder="pylate-index",
    index_name="index",
)

Retrieving top-k documents for queries

Once the documents are indexed, you can retrieve the top-k most relevant documents for a given set of queries. To do so, initialize the ColBERT retriever with the index you want to search in, encode the queries and then retrieve the top-k documents to get the top matches ids and relevance scores:

# Step 1: Initialize the ColBERT retriever
retriever = retrieve.ColBERT(index=index)

# Step 2: Encode the queries
queries_embeddings = model.encode(
    ["query for document 3", "query for document 1"],
    batch_size=32,
    is_query=True,  #  # Ensure that it is set to False to indicate that these are queries
    show_progress_bar=True,
)

# Step 3: Retrieve top-k documents
scores = retriever.retrieve(
    queries_embeddings=queries_embeddings,
    k=10,  # Retrieve the top 10 matches for each query
)

Reranking

If you only want to use the ColBERT model to perform reranking on top of your first-stage retrieval pipeline without building an index, you can simply use rank function and pass the queries and documents to rerank:

from pylate import rank, models

queries = [
    "query A",
    "query B",
]

documents = [
    ["document A", "document B"],
    ["document 1", "document C", "document B"],
]

documents_ids = [
    [1, 2],
    [1, 3, 2],
]

model = models.ColBERT(
    model_name_or_path=ayushexel/colbert-MiniLM-L6-H384-uncased-1-neg-1-epoch-gooaq-1995000,
)

queries_embeddings = model.encode(
    queries,
    is_query=True,
)

documents_embeddings = model.encode(
    documents,
    is_query=False,
)

reranked_documents = rank.rerank(
    documents_ids=documents_ids,
    queries_embeddings=queries_embeddings,
    documents_embeddings=documents_embeddings,
)

Evaluation

Metrics

Col BERTTriplet

• Evaluated with pylate.evaluation.colbert_triplet.ColBERTTripletEvaluator

Metric	Value
accuracy	0.3738

Training Details

Training Dataset

Unnamed Dataset

• Size: 1,893,949 training samples
• Columns: question, answer, and negative
• Approximate statistics based on the first 1000 samples:

  	question	answer	negative
  type	string	string	string
  details	min: 9 tokens mean: 12.73 tokens max: 23 tokens	min: 14 tokens mean: 31.78 tokens max: 32 tokens	min: 16 tokens mean: 31.7 tokens max: 32 tokens

• Samples:

  question	answer	negative
  how do i import photos from iphone onto mac?	['Open the Photos app.', 'Connect your iPhone to Mac using a USB cable.', 'In the upper menu of the Photos app, choose Import.', 'Here you will see all the photos your iPhone has.', 'To import all photos, click Import all new photos on the upper-right corner of the window.']	Import to your Mac Connect your iPhone, iPad, or iPod touch to your Mac with a USB cable. Open the Photos app. The Photos app shows an Import screen with all the photos and videos that are on your connected device. If the Import screen doesn't automatically appear, click the device's name in the Photos sidebar.
  what are hyperplastic colon polyps?	A hyperplastic polyp is a growth of extra cells that projects out from tissues inside your body. They occur in areas where your body has repaired damaged tissue, especially along your digestive tract. Hyperplastic colorectal polyps happen in your colon, the lining of your large intestine.	During the colonoscopy, it's hard to differentiate between the benign hyperplastic and the more worrisome adenomatous polyp. Polyps appear as lumps inside the colon. Some are flat and others hang down from a stalk. Each polyp is biopsied and tissue from the polyp is sent to a lab and tested for cancer.
  what are the flaws of the electoral college quizlet?	['the winner of the popular vote is not guaranteed the presidency. ... ', 'electors are not required to vote in accord with the popular vote. ... ', 'any election might have to be decided in the HOR. ... ', 'small states are overrepresented- they have more electoral votes per a smaller amount of people than larger states.']	In other U.S. elections, candidates are elected directly by popular vote. But the president and vice president are not elected directly by citizens. Instead, they're chosen by “electors” through a process called the Electoral College. ... It was a compromise between a popular vote by citizens and a vote in Congress.

• Loss: pylate.losses.contrastive.Contrastive

Evaluation Dataset

Unnamed Dataset

• Size: 5,000 evaluation samples
• Columns: question, answer, and negative_1
• Approximate statistics based on the first 1000 samples:

  	question	answer	negative_1
  type	string	string	string
  details	min: 9 tokens mean: 12.84 tokens max: 23 tokens	min: 18 tokens mean: 31.77 tokens max: 32 tokens	min: 15 tokens mean: 31.47 tokens max: 32 tokens

• Samples:

  question	answer	negative_1
  1 cup how many grams of flour?	A cup of all-purpose flour weighs 4 1/4 ounces or 120 grams. This chart is a quick reference for volume, ounces, and grams equivalencies for common ingredients.	Convert 25 grams or g of flour to cups. 25 grams flour equals 1/4 cup.
  is lasker rink owned by trump?	Lasker Rink was announced in 1962 and completed in 1966. It has been operated by The Trump Organization since 1987. In 2018, the city announced that the rink would be closed and rebuilt between 2021 and 2024.	Lasker Rink was announced in 1962 and completed in 1966. It has been operated by The Trump Organization since 1987. In 2018, the city announced that the rink would be closed and rebuilt between 2021 and 2024.
  how many litres of water to drink a day for weight loss?	Bottom Line: According to the studies, 1–2 liters of water per day is enough to assist with weight loss, especially when consumed before meals.	Based on the studies, drinking 1-2 liters of water per day should be sufficient to help with weight loss. Here's how much water you should drink, in different measurements: Liters: 1–2.

• Loss: pylate.losses.contrastive.Contrastive

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 128
• per_device_eval_batch_size: 128
• learning_rate: 3e-06
• num_train_epochs: 1
• warmup_ratio: 0.1
• seed: 12
• bf16: True
• dataloader_num_workers: 12
• load_best_model_at_end: True

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 128
• per_device_eval_batch_size: 128
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 1
• eval_accumulation_steps: None
• torch_empty_cache_steps: None
• learning_rate: 3e-06
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1.0
• num_train_epochs: 1
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.1
• warmup_steps: 0
• log_level: passive
• log_level_replica: warning
• log_on_each_node: True
• logging_nan_inf_filter: True
• save_safetensors: True
• save_on_each_node: False
• save_only_model: False
• restore_callback_states_from_checkpoint: False
• no_cuda: False
• use_cpu: False
• use_mps_device: False
• seed: 12
• data_seed: None
• jit_mode_eval: False
• use_ipex: False
• bf16: True
• fp16: False
• fp16_opt_level: O1
• half_precision_backend: auto
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• local_rank: 0
• ddp_backend: None
• tpu_num_cores: None
• tpu_metrics_debug: False
• debug: []
• dataloader_drop_last: False
• dataloader_num_workers: 12
• dataloader_prefetch_factor: None
• past_index: -1
• disable_tqdm: False
• remove_unused_columns: True
• label_names: None
• load_best_model_at_end: True
• ignore_data_skip: False
• fsdp: []
• fsdp_min_num_params: 0
• fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
• fsdp_transformer_layer_cls_to_wrap: None
• accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
• deepspeed: None
• label_smoothing_factor: 0.0
• optim: adamw_torch
• optim_args: None
• adafactor: False
• group_by_length: False
• length_column_name: length
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• skip_memory_metrics: True
• use_legacy_prediction_loop: False
• push_to_hub: False
• resume_from_checkpoint: None
• hub_model_id: None
• hub_strategy: every_save
• hub_private_repo: None
• hub_always_push: False
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• include_inputs_for_metrics: False
• include_for_metrics: []
• eval_do_concat_batches: True
• fp16_backend: auto
• push_to_hub_model_id: None
• push_to_hub_organization: None
• mp_parameters:
• auto_find_batch_size: False
• full_determinism: False
• torchdynamo: None
• ray_scope: last
• ddp_timeout: 1800
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• dispatch_batches: None
• split_batches: None
• include_tokens_per_second: False
• include_num_input_tokens_seen: False
• neftune_noise_alpha: None
• optim_target_modules: None
• batch_eval_metrics: False
• eval_on_start: False
• use_liger_kernel: False
• eval_use_gather_object: False
• average_tokens_across_devices: False
• prompts: None
• batch_sampler: batch_sampler
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	accuracy
0	0	-	0.3738
0.0001	1	9.8144	-
0.0135	200	8.6046	-
0.0270	400	6.3812	-
0.0405	600	4.0823	-
0.0541	800	2.3103	-
0.0676	1000	1.7525	-
0.0811	1200	1.4658	-
0.0946	1400	1.2898	-
0.1081	1600	1.1659	-
0.1216	1800	1.0575	-
0.1352	2000	1.0146	-
0.1487	2200	0.9502	-
0.1622	2400	0.9233	-
0.1757	2600	0.8957	-
0.1892	2800	0.8514	-
0.2027	3000	0.8499	-
0.2163	3200	0.8311	-
0.2298	3400	0.8007	-
0.2433	3600	0.787	-
0.2568	3800	0.7648	-
0.2703	4000	0.7538	-
0.2838	4200	0.7373	-
0.2974	4400	0.732	-
0.3109	4600	0.7335	-
0.3244	4800	0.7084	-
0.3379	5000	0.7109	-
0.3514	5200	0.7091	-
0.3649	5400	0.691	-
0.3785	5600	0.6814	-
0.3920	5800	0.6817	-
0.4055	6000	0.6694	-
0.4190	6200	0.6602	-
0.4325	6400	0.6594	-
0.4460	6600	0.6526	-
0.4596	6800	0.6433	-
0.4731	7000	0.6378	-
0.4866	7200	0.6362	-
0.5001	7400	0.6273	-
0.5136	7600	0.6293	-
0.5271	7800	0.6198	-
0.5407	8000	0.6166	-
0.5542	8200	0.6194	-
0.5677	8400	0.618	-
0.5812	8600	0.6109	-
0.5947	8800	0.6145	-
0.6082	9000	0.598	-
0.6217	9200	0.5982	-
0.6353	9400	0.5989	-
0.6488	9600	0.5926	-
0.6623	9800	0.5956	-
0.6758	10000	0.597	-
0.6893	10200	0.5803	-
0.7028	10400	0.5889	-
0.7164	10600	0.5907	-
0.7299	10800	0.5904	-
0.7434	11000	0.5857	-
0.7569	11200	0.5825	-
0.7704	11400	0.5825	-
0.7839	11600	0.5786	-
0.7975	11800	0.5797	-
0.8110	12000	0.5746	-
0.8245	12200	0.577	-
0.8380	12400	0.5765	-
0.8515	12600	0.5803	-
0.8650	12800	0.5671	-
0.8786	13000	0.5716	-
0.8921	13200	0.5822	-
0.9056	13400	0.5806	-
0.9191	13600	0.5734	-
0.9326	13800	0.578	-
0.9461	14000	0.569	-
0.9597	14200	0.5637	-
0.9732	14400	0.5777	-
0.9867	14600	0.5653	-

Framework Versions

• Python: 3.11.0
• Sentence Transformers: 4.0.1
• PyLate: 1.1.7
• Transformers: 4.48.2
• PyTorch: 2.6.0+cu124
• Accelerate: 1.6.0
• Datasets: 3.5.0
• Tokenizers: 0.21.1

Citation

BibTeX

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084"
}

PyLate

@misc{PyLate,
title={PyLate: Flexible Training and Retrieval for Late Interaction Models},
author={Chaffin, Antoine and Sourty, Raphaël},
url={https://github.com/lightonai/pylate},
year={2024}
}