hard_eval_outputs/hard_eval_generation.json

{
  "model": "/workspace/models/Qwen3-32B",
  "hard_scale": 1.0,
  "total": 730,
  "counts": {
    "eval_real_hard_abstain": 185,
    "eval_real_hard_seen": 272,
    "eval_real_hard_missing": 134,
    "eval_real_hard_parallel": 139
  },
  "by_scenario": {
    "hard_negative_consumer": 48,
    "hard_negative_revenue": 42,
    "hard_negative_coverage": 48,
    "hard_irrelevance": 47,
    "hard_rare_location": 84,
    "hard_disambiguation": 88,
    "hard_no_train_overlap": 100,
    "hard_missing_multi": 80,
    "hard_missing_natural": 40,
    "hard_missing_rare_slot": 14,
    "hard_parallel_implicit": 139
  }
}