license: mit
language:
- en
task_categories:
- visual-question-answering
tags:
- medical
- pathology
- pathvqa
PathVQA — grouped QA TARs
This unofficial preparation derives from the public flaviagiammarino/path-vqa mirror, pinned to revision 1685832883334b5bb5beaf4e4b333fdeecaa4ad9. The mirror documents that it used the authors' February 15, 2023 release. All downloaded Parquet SHA256 checksums were verified against the Hub metadata.
Credit: Xuehai He, Yichen Zhang, Luntian Mou, Eric Xing, and Pengtao Xie. PathVQA: 30000+ Questions for Medical Visual Question Answering, 2020. Paper. Original challenge site.
The upstream mirror declares the MIT license. Its notice about third-party image rights is retained: the dataset contains material from Textbook of Pathology, Basic Pathology, and the Pathology Education Informational Resource (PEIR); copyrights in those images and captions belong to their respective publishers, authors, and owners. The mirror's MIT declaration does not transfer ownership of those materials.
Version and splits
This release contains 32,632 QA pairs, not the original paper's 32,799. The mirror already removed duplicate image/question/answer triplets and normalized text case and spacing. This preparation preserves the mirror's questions and answers verbatim, without further QA deduplication.
| Split | Unique image rows | QA pairs | Yes/no QA | Other open-text QA |
|---|---|---|---|---|
| train (original train + validation) | 3,430 | 25,913 | 12,876 | 13,037 |
| test | 858 | 6,719 | 3,362 | 3,357 |
Original mirror image counts are 2,599 train, 832 validation, and 858 test. One image has identical decoded pixels in both train and validation; it is grouped into one row with all its QA preserved. No exact trainval/test image overlap was found, and no QA were removed. Test is preserved in full.
The mirror uses per-QA filenames rather than original image identifiers. Images are grouped by SHA256 of decoded RGB pixels plus dimensions, and output names use that hash. Source and exported image hashes are disjoint across trainval/test. This does not certify patient-level separation or detect visually similar, cropped, or re-encoded images whose decoded pixels differ.
Processing and format
Images are JPEG quality 95 at original dimensions, with no padding or resizing. CMYK and other unsupported JPEG modes are converted to RGB. One JSONL row contains all QA for one image, with human and gpt conversation roles. GPT values are <answer>original mirror answer</answer>. PathVQA supplies no bbox annotations or reasoning traces, so bbox_annotations is empty and no boxes or thinking were generated. Existing arrows or labels drawn into source figures remain part of the image; they are not separate box coordinates.
Each TAR contains exactly one matching JSONL and an image/ directory. Shards have at most 10,000 image rows, with train and test sharded independently. image values are filenames relative to image/, matching GEMeX packaging, and tar_name names the TAR. Standalone JSONL copies are provided beside the archives. Extract each shard into its own directory.
partition is a compatibility field (LLaVA-OneVision-2-Spatial), not a claim that PathVQA belongs to that upstream dataset. Preparation and packaging reports record counts and SHA256 checksums. Every archived file was checked against the prepared source before upload.