Please read and agree to the following terms before accessing CleanSlide: (1) This resource will be used for scientific research only, and not for any commercial or clinical purpose. (2) CleanSlide will be cited in any publication that uses this data.

CleanSlide — pan-cancer whole-slide vision–language benchmark

Official dataset for CleanSlide, the largest public and cleanest whole-slide-image (WSI) vision–language multiple-choice benchmark on TCGA. This repository is self-contained: pre-computed features, the questions, the splits, and a slide-download manifest are all here. Full method, code, and figures live in the CleanSlide GitHub repository.

Abstract

Pathology vision–language models are typically evaluated on TCGA whole-slide VQA, but those benchmarks leak information at two levels — patient (slides from one case split across train and test) and tissue-source-site (staining/scanner batch shared across cases) — and their questions are often answerable without ever looking at the slide. CleanSlide is a contamination-controlled benchmark of 148,654 four-option questions over 9,985 TCGA diagnostic (FFPE) WSIs spanning 32 solid cancer types. Its train / val / test folds are patient- and site (TSS)-disjoint (0 overlap), every question is audited on four cleanliness dimensions with the blind (image-free) baseline published rather than hidden, and slide features come from H-optimus-0 (Apache-2.0) — an encoder not trained on TCGA — so the released features are redistributable and encoder-uncontaminated.

Dataset structure

features/
├── train/<slide_id>.h5        # H-optimus-0 patch features, one HDF5 file per slide
├── val/<slide_id>.h5
└── test/<slide_id>.h5
questions/
├── tier1/{train,val,test}.jsonl   # Tier 1 — reused & cleaned public TCGA-WSI MCQ
└── tier2/{train,val,test}.jsonl   # Tier 2 — self-generated, text-only MCQ
slides/
├── slides_splits.csv          # slide_id, patient, tss, tumor, fold, split
└── slide_gdc_manifest.csv     # GDC manifest to download the raw diagnostic WSIs

~480 GB of features (fp16), one .h5 per slide, already partitioned by split.
148,654 questions total, separated by tier (reuse-and-clean vs self-generated) and by split.
slides/ provides the slide → fold/split table and a ready-to-use GDC download manifest for the raw WSIs.

Feature file format (`.h5`)

dataset	shape	dtype	meaning
`features`	`[N, 1536]`	float16	H-optimus-0 embedding per kept patch
`coords`	`[N, 2]`	int32	level-0 (x, y) pixel coordinate of each patch — the encoder-agnostic canonical layer: any future encoder can be re-run from these coordinates without re-segmenting/re-tiling

Per-file attributes: encoder, mpp_x, target_mpp (0.5 µm/px = 20×), patch_size (224), patch_size_lvl0, level_used, n_patches, tissue_frac_thresh (0.10), mpp_source. Extraction follows the H-optimus-0 recipe with CLAM tissue detection and the Prov-GigaPath valid-patch rule (no stain normalization — the site signature is controlled by the split, not the pixels). Details on GitHub.

Usage

Step 1 — download

from huggingface_hub import snapshot_download

# Everything (the features are ~480 GB):
snapshot_download("eric-1w/CleanSlide-features", repo_type="dataset", local_dir="cleanslide")

# Or just the questions + splits (small), skipping the large feature files:
snapshot_download("eric-1w/CleanSlide-features", repo_type="dataset", local_dir="cleanslide",
                  allow_patterns=["questions/*", "slides/*"])

Step 2 — pair a question with its slide features

The slide and split fields of each question point to features/<split>/<slide_id>.h5.

import json, h5py

q = [json.loads(l) for l in open("cleanslide/questions/tier1/test.jsonl")][0]
print(q["question"])
print(q["options"], "->", q["answer"])

with h5py.File(f"cleanslide/features/{q['split']}/{q['slide']}.h5", "r") as f:
    feats  = f["features"][:]   # (N, 1536) float16
    coords = f["coords"][:]     # (N, 2) int32, level-0 pixels

Step 3 — (optional) get the raw slides

The raw WSIs are not redistributed here. To download them, use the included GDC manifest with the GDC Data Transfer Tool:

gdc-client download -m slides/slide_gdc_manifest.csv

You can then re-extract features with the CleanSlide pipeline (code/extract/extract_features.py on GitHub).

A question record

Each line of a questions/*/*.jsonl file is one four-option question:

{
  "source": "SlideBench",
  "slide": "TCGA-08-0244-01Z-00-DX1",
  "patient": "TCGA-08-0244",
  "tss": "08",
  "tumor": "GBM",
  "question": "Examine the cellular morphology in the provided whole slide image of glioblastoma. Which cytoplasmic feature is typically observed?",
  "options": ["Mucin production", "Clear cytoplasm", "Fibrillary background", "Granulomatous changes"],
  "answer": "C",
  "answer_type": "mcq",
  "task": "Microscopy",
  "fold": 3,
  "split": "test"
}

field	meaning
`source`	origin set — `WSI-Bench`, `SlideBench`, `WSI-VQA`, `CleanSlide-labels`, or `CleanSlide-reports`
`slide` / `patient` / `tss`	TCGA slide barcode, patient, and tissue-source-site
`tumor`	TCGA study (cancer-type) code, e.g. `GBM`
`question` / `options` / `answer`	stem, four options, correct letter (`A`–`D`)
`task`	question category (Diagnosis, Microscopy, Subtype, Grading, …)
`fold` / `split`	disjoint fold id and `train` / `val` / `test`

Citation

If you use CleanSlide, please cite:

@misc{zhang2026cleanslide,
  title        = {CleanSlide: A Leakage-Audited and Shortcut-Controlled Benchmark for Whole-Slide Vision–Language Models},
  author       = {Zhang, Wenhao and others},
  year         = {2026},
  note         = {Manuscript in preparation},
  howpublished = {\url{https://github.com/wenhaozhang0066/CleanSlide}}
}

License & provenance

Questions and derived features: CC-BY-NC 4.0 (research / non-commercial use only). Underlying TCGA slides follow the NIH/GDC data policy — raw slides are not redistributed here. Encoder: H-optimus-0 (Apache-2.0), not trained on TCGA.

Downloads last month: 29

Total file size:

482 GB

You need to agree to share your contact information to access this dataset