{
  "domain": "life-health",
  "image": "python:3.12-slim",
  "data": "Public: the ODELIA training split, with the fold assignments fixed in folds.csv. The held-out folds are scored on the runner only.",
  "benchmark": {
    "task": "Breast MRI lesion classification",
    "metric": {
      "name": "AUROC",
      "better": "higher"
    },
    "cost": "GPU-hours",
    "subtasks": "folds",
    "guardrails": [
      "Each fold’s model trains on the other folds only",
      "No pretrained weights other than ImageNet",
      "At most 40 GPU-hours per scored run on the reference runner"
    ],
    "aggregate": "mean",
    "baseline": {
      "name": "3D ResNet-18",
      "score": 0.876,
      "cost": 4.1,
      "subtasks": {
        "fold 1": 0.8893,
        "fold 2": 0.865,
        "fold 3": 0.8947,
        "fold 4": 0.8792,
        "fold 5": 0.8584,
        "fold 6": 0.8901,
        "fold 7": 0.8553
      }
    }
  }
}