verification_summary.json

{
  "n_test": 100,
  "valid_json_rate": 1.0,
  "nonempty_rate": 1.0,
  "exact_set_match": 0.16,
  "mean_set_f1": 0.16,
  "average_codes": 1.0,
  "model_name": "qwen25_7b",
  "base_model": "Qwen/Qwen2.5-7B-Instruct",
  "adapter": "artifacts/qwen25_7b",
  "task_mode": "single_code",
  "thresholds": {
    "min_valid_json_rate": 0.9,
    "min_nonempty_rate": 0.95,
    "max_average_codes": 3.0
  },
  "passed": true
}