mbzuai-ai4bio/MLE-Bench-Bio8-Public

MLE-Bench — Biomedical Subset (8 tasks), public side

This repository holds only what an agent is allowed to see. The held-out labels live in a separate dataset, mbzuai-ai4bio/MLE-Bench-Bio8-Private, so a training job pointed at this repo cannot reach them even by accident.

The biomedical tasks of MLE-Bench (OpenAI), prepared with MLE-Bench's own prepare.py and split into public/ and private/ following the BioXArena convention.

⚠️ Read the licensing section before using or resharing this data.

Why 8 and not 12

The biomedical slice of MLE-Bench's 75-competition split is 12 tasks, but 4 of them — histopathologic-cancer-detection, osic-pulmonary-fibrosis-progression, rsna-miccai-brain-tumor-radiogenomic-classification and uw-madison-gi-tract-image-segmentation — also appear in BioMLBench. To avoid double-counting in cross-benchmark comparisons, they are published once, in mbzuai-ai4bio/BioMLBench-24, and this repo holds the 8 that are unique to MLE-Bench.

Tasks

Task	Metric	Higher better	MLE-Bench tier	Top-1 human
bms-molecular-translation	Levenshtein distance	✗	high	0.51
ranzcr-clip-catheter-line-classification	AUC-ROC	✓	low	0.97671
rsna-2022-cervical-spine-fracture-detection	weighted multi-label log loss	✗	high	0.2047
rsna-breast-cancer-detection	probabilistic F1	✓	high	0.55
siim-covid19-detection	mean average precision	✓	high	0.635
siim-isic-melanoma-classification	AUC-ROC	✓	low	0.9490
stanford-covid-vaccine	multi-class log loss	✗	high	0.34198
vinbigdata-chest-xray-abnormalities-detection	mAP @ IoU>0.4	✓	high	0.314

"Top-1 human" is the best score on the competition's private leaderboard, shipped as leaderboard.csv in the MLE-Bench repo and used there to compute medal thresholds.

Layout

<task>/public/
├── description.md
├── sample_submission.csv
├── train.csv / train.json / train_images/ ...
└── test.csv / test_images/ ...     # labels removed

The matching -Private repo contributes <task>/private/, so downloading both into one directory rebuilds the full layout.

Six of the eight tasks are archived as <task>.tar.zst rather than loose files — together they hold over 3.3 M image files, far past what the Hub handles comfortably. Extract with:

hf download mbzuai-ai4bio/MLE-Bench-Bio8 <task>.tar.zst --repo-type dataset --local-dir .
sha256sum -c SHA256SUMS --ignore-missing
tar --use-compress-program=unzstd -xf <task>.tar.zst

stanford-covid-vaccine (5 files) and any other small task stay as loose files so the Hub can preview them.

Splits

MLE-Bench does not reuse Kaggle's original train/test split — the public test labels were never released. It re-splits the labelled training set into a new train/test pair, so private/ holds real labels and grading is reproducible offline. Row counts and the exact split logic are in each competition's prepare.py upstream.

There is no validation split. Carve one out of train yourself.

Grading

Use MLE-Bench's own graders — each task has a bespoke metric implementation:

git clone https://github.com/openai/mle-bench && cd mle-bench && pip install -e .
mlebench grade-sample submission.csv <task> --data-dir /path/to/extracted

Point --data-dir at a tree where <task>/prepared/{public,private} matches this repo's <task>/{public,private}.

Licensing — please read

This is Kaggle competition data. MLE-Bench itself deliberately ships only download scripts and no data, because most competition rules restrict redistribution; several of these competitions (RSNA, SIIM, VinBigData) additionally restrict sharing outside the team that accepted the rules.

This repository is private and intended for internal MBZUAI AI4Bio research use. Do not make it public and do not reshare the archives without checking each competition's rules page first. If you only need a couple of tasks, downloading them yourself with mlebench prepare -c <task> and your own Kaggle credentials is the cleaner path.

MLE-Bench's own code is MIT-licensed; that license covers the harness, not the data.

@article{chan2024mlebench,
  title  = {MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering},
  author = {Chan, Jun Shern and others},
  year   = {2024},
  eprint = {2410.07095},
  archivePrefix = {arXiv}
}

Packaged by MBZUAI AI4Bio.