---
license: pddl
---

# PubMed Central Figures Dataset
 
This dataset contains image-text pairs extracted from figures from papers in the [PubMed Central](https://www.ncbi.nlm.nih.gov/pmc/) repository. The dataset can be used to train [CLIP](https://arxiv.org/abs/2103.00020) models.

This repo contains contains a [Parquet](https://parquet.apache.org/) file containing the metadata of a [WebDataset](https://github.com/webdataset/webdataset) in [img2dataset](https://github.com/rom1504/img2dataset) format. The images themselves are not distributed and need to be retrieved. Note that the images cannot be retrieved by an HTTP URL, so [img2dataset](https://github.com/rom1504/img2dataset) cannot be used as is to retrieve the data. Instead, the paper id (e.g. PMC7202302) and file name (e.g. gr3.jpg) are provided as identifier for each sample. The papers themselves can be downloaded from the [FTP server](https://www.ncbi.nlm.nih.gov/pmc/tools/ftp/).

Furthermore, the repo contains a NumPy file which contains the uid of all samples that are not considered duplicates to the [DataComp](https://datacomp.ai) evaluation data. This file can be used to decontaminate the dataset.