{
  "input_file": "/home/ziqiang/LLaMA-Factory/data/dataset/9_17/9.17_evaluate_data_top5_final.json",
  "models": [
    "/data/models/Qwen3-8B",
    "my_lora"
  ],
  "baseline_model": "/data/models/Qwen3-8B",
  "runs": {
    "/data/models/Qwen3-8B": {
      "output_file": "evaluation/multi_baseline_lora/result__data_models_Qwen3-8B.json",
      "summary": {
        "total_conversations": 397,
        "total_pairs": 1191,
        "pair_metrics": {
          "pair1": {
            "total": 397,
            "accuracy": 0.895,
            "precision@1": 1.0
          },
          "pair2": {
            "total": 397,
            "accuracy": 0.778,
            "precision@1": 0.872
          },
          "pair2_consider_recall": {
            "total": 332,
            "accuracy": 0.789,
            "precision@1": 0.895
          },
          "pair2_recall_subset": {
            "total": 332,
            "accuracy": 0.789,
            "precision@1": 0.895
          },
          "pair3": {
            "total": 397,
            "answer_score": 0.217
          }
        },
        "recall_metrics": {
          "total_pairs": 397,
          "recall@5_1": 332,
          "recall@5_0": 65,
          "recall_rate": 0.836
        },
        "overall_metrics": {
          "total": 1191,
          "accuracy": 0.836,
          "precision@1": 0.936,
          "answer_score": 0.217
        },
        "baseline": {
          "enabled": false,
          "is_baseline": false,
          "baseline_model": "/data/models/Qwen3-8B",
          "current_model": "my_lora"
        }
      }
    },
    "my_lora": {
      "output_file": "evaluation/multi_baseline_lora/result_my_lora.json",
      "summary": {
        "total_conversations": 397,
        "total_pairs": 1191,
        "pair_metrics": {
          "pair1": {
            "total": 397,
            "accuracy": 0.895,
            "precision@1": 1.0
          },
          "pair2": {
            "total": 397,
            "accuracy": 0.779,
            "precision@1": 0.872
          },
          "pair2_consider_recall": {
            "total": 333,
            "accuracy": 0.791,
            "precision@1": 0.895
          },
          "pair2_recall_subset": {
            "total": 333,
            "accuracy": 0.791,
            "precision@1": 0.895
          },
          "pair3": {
            "total": 397,
            "answer_score": 0.218
          }
        },
        "recall_metrics": {
          "total_pairs": 397,
          "recall@5_1": 333,
          "recall@5_0": 64,
          "recall_rate": 0.839
        },
        "overall_metrics": {
          "total": 1191,
          "accuracy": 0.837,
          "precision@1": 0.936,
          "answer_score": 0.218
        },
        "baseline": {
          "enabled": false,
          "is_baseline": false,
          "baseline_model": "/data/models/Qwen3-8B",
          "current_model": "my_lora"
        }
      }
    }
  },
  "comparison": {
    "/data/models/Qwen3-8B": {
      "overall_metrics": {
        "total": 1191,
        "accuracy": 0.836,
        "precision@1": 0.936,
        "answer_score": 0.217
      },
      "pair1": {
        "total": 397,
        "accuracy": 0.895,
        "precision@1": 1.0
      },
      "pair2": {
        "total": 397,
        "accuracy": 0.778,
        "precision@1": 0.872
      },
      "pair3": {
        "total": 397,
        "answer_score": 0.217
      }
    },
    "my_lora": {
      "overall_metrics": {
        "total": 1191,
        "accuracy": 0.837,
        "precision@1": 0.936,
        "answer_score": 0.218
      },
      "pair1": {
        "total": 397,
        "accuracy": 0.895,
        "precision@1": 1.0
      },
      "pair2": {
        "total": 397,
        "accuracy": 0.779,
        "precision@1": 0.872
      },
      "pair3": {
        "total": 397,
        "answer_score": 0.218
      }
    }
  }
}