初始化项目，由ModelHub XC社区提供模型

Model: lldois/v35_v29_cot_r3_balanced_lr6e7_ep10
Source: Original Platform

experiment_recipe.json

@@ -0,0 +1,72 @@
+{
+  "created_at": "2026-07-15 00:59:15",
+  "name": "v35_v29_cot_r3_balanced_lr6e7_ep10",
+  "base": "/home/ll/llm4rec/experiments/outputs/v29_v19_user_world_guard_lr8e7_ep018",
+  "dataset": "v35_cot_r3_balanced",
+  "dataset_manifest": {
+    "name": "v35_cot_r3_balanced",
+    "path": "/home/ll/llm4rec/experiments/data/v35_cot_r3_balanced.jsonl",
+    "records": 80000,
+    "bytes": 310914202,
+    "sha256": "4207a66e3c1e6fe035067e219a3a45243c46a9167f1b912eca13d62ec1cdf192",
+    "requested_components": {
+      "general": 24000,
+      "r0": 12000,
+      "r3_think": 18000,
+      "r3_direct": 14000,
+      "old_rec": 5000,
+      "old_user": 3500,
+      "old_item": 3500
+    },
+    "groups": {
+      "r3": 32000,
+      "old_user": 3500,
+      "general": 24000,
+      "old_rec": 5000,
+      "r0": 12000,
+      "old_item": 3500
+    },
+    "variants": {
+      "profile_goods_direct": 3190,
+      "original_competition": 5884,
+      "official_general_cot": 11903,
+      "user_array_json_strict": 607,
+      "official_general_direct": 12097,
+      "rec_no_think_direct_final": 2500,
+      "profile_live_direct": 3236,
+      "caption_to_sid_direct": 5975,
+      "sid_to_caption_direct": 3085,
+      "sid_to_caption_cot": 2940,
+      "profile_video/ad_direct": 3187,
+      "profile_video/video_direct": 4387,
+      "profile_live_cot": 4191,
+      "profile_goods_cot": 4133,
+      "profile_video/video_cot": 5547,
+      "profile_video/ad_cot": 4129,
+      "user_strict_array": 606,
+      "user_strict_logic": 511,
+      "item_short_think": 629,
+      "item_no_think_direct_final": 577,
+      "user_logic_json_strict": 462,
+      "user_extra_no_think_logic": 224
+    },
+    "routes": {
+      "no_think": 42146,
+      "think": 37854
+    },
+    "target_leakage": 0,
+    "seed": 202607152
+  },
+  "learning_rate": "6.0e-7",
+  "epochs": 1.0,
+  "stages": null,
+  "note": "Higher R3 CoT ratio with compact Persona-Expansion-Transition traces and direct-route balance.",
+  "cot_policy": "Preserve meaningful CoT for /think and pure final outputs for /no_think.",
+  "checkpoint_policy": "save_strategy=no; only the final model is retained.",
+  "script": "/home/ll/llm4rec/experiments/run_v34_v43.py",
+  "script_sha256": "fbf01773c61579cf6159f64442fc0c8902714c84ed23b488bc9a6236a154f402",
+  "configs": [
+    "/home/ll/llm4rec/experiments/configs/v35_v29_cot_r3_balanced_lr6e7_ep10.yaml"
+  ],
+  "reproduce": "/home/ll/llm4rec/demo/LLaMA-Factory/.venv/bin/python /home/ll/llm4rec/experiments/run_v34_v43.py --single v35_v29_cot_r3_balanced_lr6e7_ep10 --gpu <gpu>"
+}