MedWeave Evaluation Set

This private repository is the frozen, test-only export used by the MedWeave evaluation pipeline. The three clinical domains are deliberately separated so evaluators can load and score them independently.

Config	Clinical domain	Test questions	Included sources
eye	Eye / Ophthalmology	39,542	DDR (12,999), DeepDRiD (1,805), OCTDL (1,710), TOP (13,410), UCSD (3,392), dme_vqa (1,304), localized_medical_vqa_insegcat (4,322), ophthalvqa (600)
heart	Heart / Cardiology	15,374	CineMR-acdc (696), CineMR-mnms (1,846), CineMR-mnms2 (778), EchoVQA (12,054)
brain	Brain / Neurology	17,303	NeuroQA-BraTS-GLI (2,787), NeuroQA-BraTS-MEN (2,418), NeuroQA-CC359 (3,092), NeuroQA-HCP-YA (5,909), NeuroQA-IXI (2,933), NeuroQA-WMH (164)
Total		72,219

1. Get access and download a complete snapshot

The repository is private. Ask the owner to grant your Hugging Face account access, then authenticate locally without putting a token in source code:

pip install -U huggingface_hub datasets
hf auth login

Download the whole snapshot. This is the recommended route for multimodal evaluation because load_dataset() downloads the JSONL tables but does not guarantee that arbitrary files under images/ are present beside them.

from huggingface_hub import snapshot_download

repo_root = snapshot_download(
    repo_id="lixiangcog/MedWeave",
    repo_type="dataset",
    local_dir="MedWeave",
)
print(repo_root)

Keep the directory structure unchanged: every image value in a QA row is relative to this snapshot root.

2. Load questions and resolve images

For table-only inspection, a domain can be loaded directly:

from datasets import load_dataset

eye = load_dataset(
    "lixiangcog/MedWeave",
    "eye",
    split="test",
    token=True,
)

For actual inference, load the downloaded shards and resolve the image path against repo_root:

from pathlib import Path
from datasets import load_dataset

root = Path(repo_root)
shards = sorted(str(p) for p in (root / "data" / "eye").glob("test-*.jsonl"))
eye = load_dataset("json", data_files=shards, split="train")

for row in eye:
    image_path = root / row["image"]
    if row["image_available_on_hub"]:
        assert image_path.is_file(), image_path
    # Send row["question"], the choices when present, and image_path to the model.

To run immediately using only redistributed images:

ready_eye = eye.filter(lambda row: row["image_available_on_hub"])

The same procedure applies to heart and brain. Do not treat unavailable-image rows as text-only examples: image_required is true, so those rows become runnable only after the external images have been materialized.

3. Record schema

The normalized files under data/ use one schema across all domains:

Field	Meaning
id	Stable MedWeave evaluation ID.
question	Exact model-facing question.
response_format	multiple_choice, open, or the frozen source-specific format.
option_labels, option_texts	Ordered choices; empty for open questions.
answer, answer_letter, answer_numeric, answer_unit	Ground truth for offline scoring only; never insert these fields into the prompt.
numeric_tolerance	Accepted absolute tolerance for numerical questions, when applicable.
image	Snapshot-relative path of the exact visual input.
image_sha256	SHA-256 of that visual input. This is the authoritative identity check.
image_available_on_hub	Whether the image is redistributed in this repository.
source_dataset, source_subset, source_split, source_item_id	Upstream provenance and source-level mapping.
specialty, domain, modality, category	Domain and task metadata for stratified reporting.

Use evaluation_manifests/<domain>/test-*.jsonl instead when reproducing the original project runner. Those records preserve runner-native fields while replacing machine-local paths with repository-relative paths.

4. Image coverage

This private repository includes 8,839 license-verified visual assets (4.91 GiB). Every included file was matched to the frozen experiment input by SHA-256 before upload.

Config	Questions with image included	Questions requiring upstream image access	Included unique assets
eye	26,132	13,410 (TOP)	6,557
heart	12,054	3,320 (CineMR)	2,282
brain	0	17,303 (NeuroQA source images)	0

TOP, CineMR (ACDC/M&Ms/M&Ms-2), and NeuroQA source images are not redistributed. Private access to this repository does not replace their upstream access conditions. See SOURCE_ACCESS.md for the official acquisition routes, the processing boundary, and source-specific instructions.

5. Understand and use the external-image mapping

image_manifests/all.jsonl contains exactly one record per required visual asset. The important fields are:

{
  "expected_path": "images/brain/<sha256>.png",
  "sha256": "<sha256>",
  "source_datasets": ["NeuroQA-HCP-YA"],
  "source_references": ["source-data://brain/images/cases/0791/model_input.png"],
  "available_on_hub": false,
  "distribution_status": "external_source_access_required"
}

The joins are deterministic:

QA row.image_sha256
        = image_manifests/all.jsonl.sha256
        = SHA-256(bytes of the prepared image)

QA row.image
        = image_manifests/all.jsonl.expected_path

source_references is an audit reference, not a public download URL. Strip the source-data:// prefix to see the source-cache-relative location used during dataset construction. For TOP it points to an upstream JPEG. For CineMR and NeuroQA it points to a derived 2-D model input (a rendered CMR panel or model_input.png), so downloading raw NIfTI volumes alone is not enough to satisfy the hash.

The repository also includes a ready-to-use image_manifests/external.csv with one row per unavailable asset and the number of affected questions. Regenerate or audit it with:

python scripts/report_image_coverage.py \
  --repo-root . \
  --csv external_image_mapping.csv

After legally obtaining and, where required, preprocessing the upstream data, place all candidate files under one or more search roots and run:

python scripts/materialize_images.py \
  --manifest image_manifests/all.jsonl \
  --output-root . \
  --search-root /path/to/TOP \
  --search-root /path/to/rendered/cinemr \
  --search-root /path/to/preprocessed/neuroqa \
  --copy

The script does not download data and does not match by filename. It hashes only the user-provided files, accepts exact SHA-256 matches, and copies (or, without --copy, symlinks) them to the required images/eye, images/heart, and images/brain paths. Existing correct files are retained; non-matching targets are never overwritten. A machine-readable missing_images.jsonl is written when any asset remains unresolved.

Run the command again after adding another source. Completion is reached only when it reports still_missing: 0. This hash gate prevents similarly named subjects, alternate preprocessing, masks, or annotations from silently entering the evaluation.

6. Evaluation alignment

Concatenate the lexicographically ordered shards for a domain. Sharding does not reorder or alter records. Keep the frozen question list, model, prompt, decoding budget, and scoring protocol unchanged when comparing methods. Do not mix development examples into the test files and do not expose any ground-truth field to the model.

The release-level checks are in AUDIT.json, image-upload checks are in IMAGE_UPLOAD_AUDIT.json, and file digests are in SHA256SUMS. Image provenance and redistribution status are documented in ATTRIBUTION.md.

License

This is a multi-source release. Questions, annotations, and images remain governed by their upstream licenses and data-use agreements; the repository's private status does not supersede those terms. No blanket license is granted over the complete collection.