初始化项目，由ModelHub XC社区提供模型

Model: lldois/v36_v29_general_guard_lr4e7_ep12
Source: Original Platform

experiment_recipe.json

@@ -0,0 +1,72 @@
+{
+  "created_at": "2026-07-15 00:59:15",
+  "name": "v36_v29_general_guard_lr4e7_ep12",
+  "base": "/home/ll/llm4rec/experiments/outputs/v29_v19_user_world_guard_lr8e7_ep018",
+  "dataset": "v36_general_guard",
+  "dataset_manifest": {
+    "name": "v36_general_guard",
+    "path": "/home/ll/llm4rec/experiments/data/v36_general_guard.jsonl",
+    "records": 80000,
+    "bytes": 322603609,
+    "sha256": "8aa50ef29b36de4764d1255110a3d1eba6c62e3eaa22636cd92f351cb59d72c2",
+    "requested_components": {
+      "general": 36000,
+      "r0": 12000,
+      "r3_think": 7000,
+      "r3_direct": 10000,
+      "old_rec": 6000,
+      "old_user": 4500,
+      "old_item": 4500
+    },
+    "groups": {
+      "general": 36000,
+      "r0": 12000,
+      "r3": 17000,
+      "old_item": 4500,
+      "old_user": 4500,
+      "old_rec": 6000
+    },
+    "variants": {
+      "official_general_cot": 17998,
+      "caption_to_sid_direct": 5982,
+      "profile_video/ad_cot": 1620,
+      "profile_video/video_cot": 2159,
+      "original_competition": 7355,
+      "profile_goods_direct": 2199,
+      "official_general_direct": 18002,
+      "profile_live_direct": 2330,
+      "profile_video/ad_direct": 2329,
+      "user_strict_logic": 645,
+      "sid_to_caption_cot": 2964,
+      "rec_no_think_direct_final": 2970,
+      "profile_video/video_direct": 3142,
+      "profile_live_cot": 1623,
+      "user_strict_array": 775,
+      "user_array_json_strict": 710,
+      "item_short_think": 779,
+      "sid_to_caption_direct": 3054,
+      "profile_goods_cot": 1598,
+      "user_extra_no_think_logic": 294,
+      "item_no_think_direct_final": 819,
+      "user_logic_json_strict": 653
+    },
+    "routes": {
+      "think": 34086,
+      "no_think": 45914
+    },
+    "target_leakage": 0,
+    "seed": 202607153
+  },
+  "learning_rate": "4.0e-7",
+  "epochs": 1.2,
+  "stages": null,
+  "note": "General-heavy replay explicitly targeting the world-score forgetting seen in v31-v33.",
+  "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/v36_v29_general_guard_lr4e7_ep12.yaml"
+  ],
+  "reproduce": "/home/ll/llm4rec/demo/LLaMA-Factory/.venv/bin/python /home/ll/llm4rec/experiments/run_v34_v43.py --single v36_v29_general_guard_lr4e7_ep12 --gpu <gpu>"
+}