# PMC-OA Dataset

**News: We have released the PMC-OA dataset. You can choose the subset specifically.**

**P.S.** There's something wrong with the huggingface dataset viewer when the dataset scale gets large.
So we sample a subset of it to visualize it directly on web. Click [PMC-OA-Demo](https://huggingface.co/datasets/axiong/pmc_oa_demo) to view it.



[中文文档](./README.zh.md)

- [PMC-OA Dataset](#pmc-oa-dataset)
  - [Model Zoo](#model-zoo)
  - [Daraset Structure](#daraset-structure)
  - [Sample](#sample)


## Model Zoo
Check it out if you want to load model pretrained on PMC-OA directly.

We plan to release more models pretrained on PMC-OA. Feel free to reach us if the model you want is not included in model zoo for now.
Also, we express our thanks to the help from the community.

| Model | Link | Provider |
| --- | --- | --- |
| ViT-L-14 | https://huggingface.co/ryanyip7777/pmc_vit_l_14 | @ryanyip7777 |


## Daraset Structure

**PMC-OA** (seperated images, separated caption).
- `images.zip`: images folder
- `pmc_oa.jsonl`: dataset file of pmc-oa
- `pmc_oa_beta.jsonl`: dataset file of pmc-oa-beta

~~- `train.jsonl`: metafile of train set~~
~~- `valid.jsonl`: metafile of valid set~~
~~- `test.jsonl`: metafile of test set~~

The difference between PMC-OA & PMC-OA-Beta lies in the methods of processing captions.
In PMC-OA, we utilize ChatGPT to help us divide compound captions into seperate ones.
While PMC-OA-Beta keeps all the compound ones without division.


## Sample

A row in `pmc_oa.jsonl` is shown bellow,

```python
{
    "image": "PMC212319_Fig3_4.jpg",
    "caption": "A. Real time image of the translocation of ARF1-GFP to the plasma membrane ...",
}
```

Explanation to each key

- image: path to the image
- caption: corresponding to the image