MedVLM selected training-data snapshot
This is a partial archive of actual training inputs, not the complete MedVLM training corpus or a mirror of the full upstream datasets. It contains 301,223 training records across the following 10 sources. Images are included, not just server-local paths. Counts are instruction/answer records, not unique patients.
| Source | Alignment rows | Post-training rows | Upstream license |
|---|---|---|---|
| PMC-VQA | 104,144 | 104,465 | CC BY-SA; version unspecified in source card |
| SLAKE English | 1,772 | 2,729 | CC BY 4.0 |
| NIH ChestX-ray14 | 0 | 77,919 | CC0 |
| COVID-QU-Ex | 0 | 47,155 | CC BY-SA 4.0 |
| OrganAMNIST-224 | 0 | 34,561 | CC BY 4.0 |
| OrganCMNIST-224 | 0 | 12,975 | CC BY 4.0 |
| OrganSMNIST-224 | 0 | 13,932 | CC BY 4.0 |
| PneumoniaMNIST-224 | 0 | 4,708 | CC BY 4.0 |
| SARS-CoV-2 CT | 0 | 2,479 | CC BY-NC-SA 4.0 |
| MedSeg COVID Dataset 1 | 0 | 300 | CC0 |
Contents and changes
Each subsets/<source>/train.jsonl.gz stores structured question/class/target
information, the full English instruction rendered by the actual current
training prompt registry, and the exact training answer. image_ref points to
an included TAR member or an indexed NPY array. Original mixed-manifest line
numbers and per-record hashes preserve provenance; these hashes are calculated
before portable path conversion. Image bytes and NPY arrays are preserved.
English-only refers to the instruction/answer corpus, not an OCR guarantee about
every image. The image subset has not received a new clinical annotation review.
stages identifies alignment versus post-training membership. The included
alignment records were checked against their counterparts in the actual
post-training manifest. Classification targets are JSON labels; 2D localization
targets are semantic union boxes in normalized 0–1000 XYXY coordinates, with
exclusive maximum edges. They are not instance-level detection annotations.
Masks referenced by the released localization records are included where the
manifest records a mask file or archive/member. The original native masks and
coordinate provenance are retained; only image-side task targets are used in VLM
training. NPY-backed MedMNIST images retain their original training array indices.
Restore portable training manifests
Install Python, NumPy and Pillow, download this repository after approval, then:
python restore_training.py --root /path/to/downloaded/repo --output-dir /path/to/restored
The script validates file hashes, safely extracts only image/mask TAR members,
and creates posttrain.jsonl and alignment.jsonl with locally usable paths and
the archived full instruction. It deliberately removes template_id from the
restored rows so the archived instruction is used verbatim. The frozen registry
under code/ can alternatively regenerate the prompts. This is a selected-data
reproduction package, not an exact optimizer/batch/RNG-state training resume.
Excluded pending permission or further license verification
The initial image-backed release omitted the following sources; see the text-only update below: COV-CTR,
HeAL-MedVQA, ImageCLEF VQA-Med 2019, COVID-19 Radiography, Montgomery, or CADS slice
boxes. See RELEASE_STATUS.json for source-specific reasons. CADS combines
multiple source datasets and has not passed a complete per-source redistribution
review. Public availability of a mirror is not treated as new redistribution
permission. The full original request therefore remains only partially covered.
Attribution and terms
Each subset retains its own license; this mixed collection is not assigned one
blanket license. See its NOTICE.txt, metadata.json and license_evidence/.
License texts: CC0,
CC BY 4.0,
CC BY-SA 4.0,
CC BY-NC-SA 4.0.
For PMC-VQA, consult its original source card; this archive does not invent a
license version that the source did not specify.
Upstream licenses and attribution must accompany any further use/distribution.
SARS-CoV-2 CT is non-commercial and share-alike. PMC-VQA and COVID-QU-Ex retain
their share-alike terms. The manual access gate does not revoke rights granted by
those licenses and is not a substitute for upstream permissions.
Source links: PMC-VQA, SLAKE, NIH ChestX-ray14, COVID-QU-Ex, MedMNIST, SARS-CoV-2 CT, MedSeg.
No clinical use, diagnostic reliability or patient-level independence across upstream collections is claimed. Original source limitations and potential label noise remain. Held-out validation/test rows are not added to this training-only archive. Training code templates for excluded sources may be present but contain no corresponding data payload. Existing model/checkpoint repositories are not modified by this release.
Text-only supplement — 2026-10-04
Text supplement now includes actual ImageCLEF 2019 text (5,998 records) and CADS-derived text (5,062,366 records across 32 source collections), with real textual examples and case identifiers. No additional image, mask or volume files are included. CADS Han-Seg and actual texts from COV-CTR, HeAL-MedVQA, COVID-19 Radiography and Montgomery remain withheld. Template placeholders are explicitly distinguished from actual training records.
- Downloads last month
- 31