{
  "created_at": "2026-07-12 04:25:42",
  "name": "v30_soup_v19_v21_v26_w45_35_20",
  "method": "weighted_model_soup",
  "sources": [
    {
      "path": "/home/ll/llm4rec/experiments/outputs/v19_v15_dualmode_fast_lr5e6_ep055",
      "weight": 0.45
    },
    {
      "path": "/home/ll/llm4rec/experiments/outputs/v21_v15_user_repair_rec_fast_lr4e6_ep055",
      "weight": 0.35
    },
    {
      "path": "/home/ll/llm4rec/experiments/outputs/v26_v21_product_rec_rebalance_lr22e6_ep032",
      "weight": 0.2
    }
  ],
  "source_scores": {
    "v19": {
      "total": 0.8855,
      "item": 0.2146,
      "user": [
        0.0855,
        0.0337
      ],
      "rec": [
        0.0768,
        0.102,
        0.1344,
        0.1143
      ],
      "world": 0.1242,
      "eval_time_min": 48.11
    },
    "v21": {
      "total": 0.88,
      "item": 0.2146,
      "user": [
        0.0881,
        0.0347
      ],
      "rec": [
        0.0576,
        0.1122,
        0.1358,
        0.1098
      ],
      "world": 0.1271,
      "eval_time_min": 47.61
    },
    "v26": {
      "total": 0.8804,
      "item": 0.2146,
      "user": [
        0.0895,
        0.0344
      ],
      "rec": [
        0.0768,
        0.0986,
        0.133,
        0.1116
      ],
      "world": 0.1219,
      "eval_time_min": 45.73
    }
  },
  "rationale": "Average related CoT-native checkpoints from the same architecture to combine v19 total/rec4, v21 user/world signal, and v26 user1/rec1/speed without adding inference-time cost.",
  "cot_policy": "No v7 final-only checkpoint is used. All sources are CoT-native continuations.",
  "reproduce": "demo/LLaMA-Factory/.venv/bin/python experiments/model_soup.py --name v30_soup_v19_v21_v26_w45_35_20 --overwrite",
  "runtime_seconds": 2.36
}