{
  "model": "base model",
  "data": "training/out-v0.5/test.jsonl",
  "count": 60,
  "seed": 5,
  "tasks": {
    "ai_classification": {
      "n": 4,
      "valid_json": 0.0,
      "grounded": 0.0,
      "fields_match": 0.0
    },
    "chat": {
      "n": 38,
      "valid_json": 1.0,
      "grounded": 0.9473684210526315,
      "fields_match": 0.0
    },
    "finding": {
      "n": 15,
      "valid_json": 0.0,
      "grounded": 0.0,
      "fields_match": 0.0
    },
    "summary": {
      "n": 3,
      "valid_json": 0.0,
      "grounded": 0.0,
      "fields_match": 0.0
    }
  },
  "overall": {
    "valid_json": 0.6333333333333333,
    "grounded": 0.6,
    "fields_match": 0.0
  }
}