PMC-VQA: Visual Instruction Tuning for Medical Visual Question Answering
Paper • 2305.10415 • Published • 1
Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 80, in _split_generators
raise ValueError(
...<2 lines>...
)
ValueError: The TAR archives of the dataset should be in WebDataset format, but the files in the archive don't share the same prefix or the same types.
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
PMC-VQA v1 — grouped QA TARs
This unofficial preparation uses version 1 of RadGenome/PMC-VQA, pinned to b56ae594f794867893143b337b4118a835794647. It uses train.csv, test.csv, test_clean.csv, and images.zip. Version 2 is not mixed into this release. All downloaded source files were checked against Hub checksums.
Credit: Xiaoman Zhang, Chaoyi Wu, Ziheng Zhao, Weixiong Lin, Ya Zhang, Yanfeng Wang, and Weidi Xie. PMC-VQA: Visual Instruction Tuning for Medical Visual Question Answering, 2023. Paper · Official code.
The upstream dataset card releases PMC-VQA under CC BY-SA, without specifying the license version; this derivative retains that designation rather than inventing a version. The upstream card states that source articles are CC0 or CC BY. article_licenses.csv preserves the provided article citations, PMC identifiers, licenses, and article links for 109,885 of 110,445 source papers. The supplied inventory lacks 560 papers; these are listed in attribution_report.json and article_links_missing_from_inventory.txt, without assigning inferred article-specific licenses. Image and article attribution remain with their original authors.
| Split | Image rows | QA pairs |
|---|---|---|
| train (no official validation split exists) | 115,797 | 176,948 |
| test | 33,248 | 50,000 |
The source contains 176,948 train QA and 50,000 test QA. All test QA are retained. 0 train QA belonging to 0 images were removed because they matched test image names, decoded pixels, or exported pixels. Identical decoded images within each split are grouped into one row with all associated QA. These checks do not certify patient-level separation or detect visually similar images whose decoded pixels differ.
test_clean.csv is an exact 2,000-row subset of test. Those QA have test_clean: true in qa_metadata; they are not appended as duplicate QA or mixed into train.
Images and QA
Images retain their original dimensions, without padding or resizing, and are encoded as JPEG quality 95. Unsupported JPEG color modes are converted to RGB. Source figure filenames are retained for traceability. One JSONL row contains all QA for an image. Conversation roles are human and gpt.
Questions contain the source question followed by four choices labeled A–D. Answers use <answer>['B']</answer> (the selected single label). Choices such as “Both A and B” remain one selectable option. Original answer text, original label, selected label, original image, source CSV row index (zero-based), split, and clean-test membership are retained in qa_metadata, aligned with conversation pairs.
Choice prefixes are normalized to their CSV columns. 29 malformed or inconsistent labels were repaired only when the answer text uniquely matched one choice after stripping prefixes, surrounding quotes, whitespace differences, and terminal periods. 2 unmatched answer texts with valid labels were preserved and flagged rather than clinically reinterpreted. annotation_audit.json retains every affected source row and correction. This is a formatting audit, not clinical validation of the automatically generated QA.
No bounding-box coordinates or reasoning traces are supplied by this v1 release. bbox_annotations is empty, and no boxes or thinking were generated. Existing figure panels, arrows, labels, or rectangles remain in the source images.
TAR layout
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. JSONL image values are filenames relative to image/; tar_name identifies the TAR. Standalone JSONL copies are provided beside archives. Extract each shard to its own directory.
partition is retained as a compatibility field (LLaVA-OneVision-2-Spatial), not an upstream membership claim. Preparation and packaging reports record counts and SHA256 checksums. Every archived file was checked against its prepared source, and all uploaded TAR checksums are verified remotely.
- Downloads last month
- 161