{
  "created_at": "2026-07-10 01:00:49",
  "run": {
    "name": "v23_scratch_clean_cot_guard_lr3e6_ep5",
    "dataset": "scratch_clean_cot_guard_ep5",
    "model_path": "OpenOneRec/OneReason-0.8B-pretrain-competition",
    "lr": "3.0e-6",
    "epochs": 5,
    "warmup": 0.02,
    "scheduler": "cosine",
    "seed": 202607232,
    "note": "Scratch SFT from official OneReason. More conservative 5-epoch CoT run with lower LR, heavier item replay, and lighter rec augmentation. Tests whether 3-5 epoch training helps when update size is controlled.",
    "config": "/home/ll/llm4rec/experiments/configs/v23_scratch_clean_cot_guard_lr3e6_ep5.yaml"
  },
  "dataset_manifest": {
    "name": "scratch_clean_cot_guard_ep5",
    "path": "/home/ll/llm4rec/experiments/data/scratch_clean_cot_guard_ep5.jsonl",
    "records": 60765,
    "groups": {
      "rec": 27204,
      "item": 27175,
      "user": 6386
    },
    "variants": {
      "raw": 32480,
      "item_compact_cot": 5597,
      "rec_no_think_direct_final": 3200,
      "rec_cot_pattern_clean": 4800,
      "item_no_think_direct_final": 5610,
      "item_short_think": 5584,
      "user_strict_array": 1588,
      "user_extra_no_think_logic": 602,
      "user_strict_logic": 1304
    },
    "sha256": "805b812a7cf3f9d86b8067d1d19b7fc116fb6739c75ab6aa04c16a4a13bb1419",
    "seed": 202607112
  },
  "dataset_recipe": "Scratch official-OneReason SFT: original CoT all + 4800 cleaned rec CoT + 3200 rec /no_think direct-final + strict dual user aug + 3x item-style replay. Lower LR and 5 epochs test longer training without v7 final-only bias.",
  "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.",
    "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": "a03ddf5a479db437247aed4b8000b25cb5021988029220bd818133123951aa00",
  "deadline": "2026-07-10 09:00:00 +0800",
  "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/v23_scratch_clean_cot_guard_lr3e6_ep5.yaml'"
  }
}