training_metadata.json

{
  "checkpoint": "sft_v13_latest.pt",
  "step": 1000,
  "loss": 0.1500312450490533,
  "parameters": 496998400,
  "eval_results": {
    "search_accuracy": 0.92,
    "evidence_rate": 1.0,
    "finish_rate": 1.0
  }
}