{
  "created_at": "2026-07-13 03:28:12",
  "run": {
    "name": "v32_v29_balanced_r3_draft_lr8e7_ep020",
    "dataset": "v29_balanced_r3_draft",
    "model_path": "/home/ll/llm4rec/experiments/outputs/v29_v19_user_world_guard_lr8e7_ep018",
    "lr": "8.0e-7",
    "epochs": 0.2,
    "warmup": 0.02,
    "scheduler": "cosine",
    "seed": 202607321,
    "note": "v29 continuation and balanced four-domain R3 specialist. Equalize live/ad/product/video target supervision and train concise individualized interest-evolution-task CoT plus route-correct direct answers. Goal: reduce video dominance and improve recommendation balance while retaining meaningful CoT.",
    "config": "/home/ll/llm4rec/experiments/configs/v32_v29_balanced_r3_draft_lr8e7_ep020.yaml"
  },
  "dataset_manifest": {
    "name": "v29_balanced_r3_draft",
    "path": "/home/ll/llm4rec/experiments/data/v29_balanced_r3_draft.jsonl",
    "records": 15942,
    "groups": {
      "rec": 10342,
      "user": 2400,
      "item": 3200
    },
    "variants": {
      "rec_r3_draft_evidence": 5171,
      "user_strict_logic": 892,
      "rec_no_think_direct_final": 5171,
      "item_no_think_direct_final": 1563,
      "user_strict_array": 1087,
      "user_extra_no_think_logic": 421,
      "item_short_think": 1637
    },
    "sha256": "a7e50c88eef370f6de88e1afa70995c7712afa750acc711b490d34d700d44f31",
    "seed": 202607112
  },
  "dataset_recipe": "v29 four-domain cognition specialist: equal target counts for live/ad/product/video in both individualized R3 draft-thinking and direct no-thinking routes, plus item and user guards.",
  "cot_policy": "Preserve /think reasoning supervision and do not use v7_final_only as a CoT training base. For /no_think prompts, train pure final answers without generated <think> tags. This is route-specific behavior, not global CoT removal.",
  "raw_counts": {
    "rec": 19204,
    "item": 10384,
    "user": 2892
  },
  "eval_observations": {
    "v07": "best local score so far: total=0.8978, eval_time≈47.3min; fast final outputs likely help.",
    "v15": "best CoT-preserving score so far: total=0.8778, eval_time≈70.1min; logs show repeated tokens, JSON shell errors, prompt leakage, and verbose /no_think outputs.",
    "v19": "best CoT-native continuation so far: total=0.8855, eval_time≈48.1min; user1 and rec4 improved but world dropped.",
    "v20": "v7 final-only continuation with CoT restore failed as a CoT route: total=0.8527, item fell to 0.1840; do not use v7 as future CoT base.",
    "v22": "scratch official-base 3 epoch clean CoT underperformed: total=0.8217; item/world preserved but user and rec2 are weak.",
    "v23": "scratch official-base 5 epoch low-LR guard failed badly: total=0.6990; item/user collapse suggests long scratch SFT is not viable with current data mix.",
    "v24": "v15 light repair is best among v22-v24 but still only total=0.8364; user2 improves but rec/world do not recover.",
    "v25": "v19 product-heavy repair did not beat v19: total=0.8793. Logs show heavy repeated product tokens and repeated short-think phrases; avoid this over-sampling pattern.",
    "v26": "best latest batch and fastest eval: total=0.8804, eval_time≈45.7min, best user1/rec1. It is useful as a base, but rec2/world dropped.",
    "v27": "v12 product/ad repair kept user2/world relatively better but was slow and template-heavy: total=0.8563, eval_time≈70.6min. Do not continue this exact direction.",
    "v28": "v26 no-template repair regressed to total=0.8748 and slowed to 51.7min; broad continuation from v26 did not recover world/rec balance.",
    "v29": "new best CoT-native model: total=0.9038, eval_time≈45.3min. Strong item/user1/ad/product/world, with live recommendation (0.1054) the clearest remaining gap. Logs still show occasional no-think tag leakage and a repeated generic live-reasoning sentence.",
    "v30": "three-source soup reached total=0.8846, below v29. It retained speed/world but diluted recommendation scores; do not repeat broad checkpoint averaging.",
    "v16": "CoT pattern rewrite failed: total=0.7912; logs show malformed user JSON and fragmented recommendation reasoning.",
    "v18": "low-LR mixed replay from v12 failed: total=0.8340; item/world dropped and rec outputs mixed text/itemic/think tags."
  },
  "script": "/home/ll/llm4rec/experiments/run_experiments.py",
  "script_sha256": "a5a06949ec583ca8f6e5c6de8690d3005ec1ef801b508b7286f1162f5b0c87b4",
  "deadline": "none",
  "reproduce": {
    "prepare_command": "EXPERIMENT_PREPARE_ONLY=1 python3 experiments/run_experiments.py",
    "train_command": "CUDA_VISIBLE_DEVICES=<gpu> bash -lc 'source /home/ll/llm4rec/demo/LLaMA-Factory/.venv/bin/activate && llamafactory-cli train /home/ll/llm4rec/experiments/configs/v32_v29_balanced_r3_draft_lr8e7_ep020.yaml'"
  }
}