初始化项目，由ModelHub XC社区提供模型

Model: lldois/v29_v19_user_world_guard_lr8e7_ep018
Source: Original Platform

experiment_recipe.json

@@ -0,0 +1,66 @@
+{
+  "created_at": "2026-07-12 04:25:55",
+  "run": {
+    "name": "v29_v19_user_world_guard_lr8e7_ep018",
+    "dataset": "v19_user_world_guard_no_template",
+    "model_path": "/home/ll/llm4rec/experiments/outputs/v19_v15_dualmode_fast_lr5e6_ep055",
+    "lr": "8.0e-7",
+    "epochs": 0.18,
+    "warmup": 0.02,
+    "scheduler": "cosine",
+    "seed": 202607291,
+    "note": "v19 continuation. Very small user/world guard update with strict user JSON, original cleaned CoT, and light item replay. Goal: preserve the best CoT total while nudging user2/world back up without adding template repetition.",
+    "config": "/home/ll/llm4rec/experiments/configs/v29_v19_user_world_guard_lr8e7_ep018.yaml"
+  },
+  "dataset_manifest": {
+    "name": "v19_user_world_guard_no_template",
+    "path": "/home/ll/llm4rec/experiments/data/v19_user_world_guard_no_template.jsonl",
+    "records": 18091,
+    "groups": {
+      "user": 5094,
+      "item": 7997,
+      "rec": 5000
+    },
+    "variants": {
+      "user_strict_array": 1588,
+      "item_no_think_direct_final": 2805,
+      "rec_cot_pattern_clean": 2800,
+      "raw": 1600,
+      "rec_no_think_direct_final": 2200,
+      "user_strict_logic": 1304,
+      "item_short_think": 2792,
+      "item_compact_cot": 2400,
+      "user_extra_no_think_logic": 602
+    },
+    "sha256": "d540db2637094ad95fb48e3daade2a84576def30b4e06a840f8af83ce46a45bb",
+    "seed": 202607112
+  },
+  "dataset_recipe": "v19 continuation: very small LR user/world preservation update with raw user CoT, strict JSON dual-route replay, light cleaned rec CoT, no short-template synthetic reasoning, and item replay.",
+  "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.",
+    "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": "7375c7a06ae5221c98bc2135819e0999d80e2b2098543ac9a5c56c352bd693d3",
+  "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/v29_v19_user_world_guard_lr8e7_ep018.yaml'"
+  }
+}