{
  "schema": 1,
  "artifactSha256": "5a798ebc1cad921e0ce639b2f10c4bcce492654a79d4b4ca52bd0268416f6b87",
  "privateHoldout": false,
  "benchmarks": [
    {
      "name": "IFEval",
      "datasetSource": "https://github.com/google-research/google-research/tree/master/instruction_following_eval",
      "datasetVersion": "public-2026-05",
      "config": {
        "name": "ifeval-config",
        "path": "public/benchmark-configs/ifeval-config.json",
        "sha256": "50a027f2ba423aa0398446baac1ac556fdae7395104fbb809dab29f8c78df373",
        "sizeBytes": 323
      },
      "public": true,
      "score": 1,
      "tolerance": {
        "metric": "absolute-score-delta",
        "value": 0.01,
        "reason": "Fixture tolerance for backend nondeterminism"
      }
    }
  ],
  "rawResults": [
    {
      "name": "ifeval-fixture-raw",
      "path": "raw/ifeval-fixture-raw.json",
      "sha256": "5a8e8c8a48478370ea0cc95f3ac8dd037c60f5e9ae00cfe3c49e732069db574b",
      "sizeBytes": 489
    }
  ],
  "summary": {
    "passed": true
  }
}