初始化项目，由ModelHub XC社区提供模型

Model: lldois/v38_v29_r3_aggressive_lr7e7_ep10
Source: Original Platform

experiment_recipe.json

@@ -0,0 +1,72 @@
+{
+  "created_at": "2026-07-15 00:59:15",
+  "name": "v38_v29_r3_aggressive_lr7e7_ep10",
+  "base": "/home/ll/llm4rec/experiments/outputs/v29_v19_user_world_guard_lr8e7_ep018",
+  "dataset": "v38_r3_aggressive",
+  "dataset_manifest": {
+    "name": "v38_r3_aggressive",
+    "path": "/home/ll/llm4rec/experiments/data/v38_r3_aggressive.jsonl",
+    "records": 80000,
+    "bytes": 305394286,
+    "sha256": "1b53237ca7e35d4c0019b89966b7eec83e6945b643ddc0fe86ed057d90f7b81c",
+    "requested_components": {
+      "general": 20000,
+      "r0": 8000,
+      "r3_think": 20000,
+      "r3_direct": 22000,
+      "old_rec": 4000,
+      "old_user": 3000,
+      "old_item": 3000
+    },
+    "groups": {
+      "r3": 42000,
+      "general": 20000,
+      "old_item": 3000,
+      "old_user": 3000,
+      "old_rec": 4000,
+      "r0": 8000
+    },
+    "variants": {
+      "profile_goods_direct": 5026,
+      "profile_goods_cot": 4597,
+      "official_general_cot": 10049,
+      "profile_video/video_cot": 6131,
+      "profile_video/video_direct": 6741,
+      "original_competition": 4883,
+      "profile_live_direct": 5061,
+      "user_logic_json_strict": 411,
+      "official_general_direct": 9951,
+      "user_strict_array": 503,
+      "caption_to_sid_direct": 3936,
+      "profile_live_cot": 4541,
+      "profile_video/ad_direct": 5172,
+      "rec_no_think_direct_final": 2005,
+      "profile_video/ad_cot": 4731,
+      "sid_to_caption_cot": 2069,
+      "user_extra_no_think_logic": 204,
+      "item_short_think": 516,
+      "item_no_think_direct_final": 555,
+      "sid_to_caption_direct": 1995,
+      "user_array_json_strict": 514,
+      "user_strict_logic": 409
+    },
+    "routes": {
+      "no_think": 43837,
+      "think": 36163
+    },
+    "target_leakage": 0,
+    "seed": 202607155
+  },
+  "learning_rate": "7.0e-7",
+  "epochs": 1.0,
+  "stages": null,
+  "note": "Highest R3 share, trained on real temporal holdouts from official user profiles with target leakage blocked.",
+  "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/v38_v29_r3_aggressive_lr7e7_ep10.yaml"
+  ],
+  "reproduce": "/home/ll/llm4rec/demo/LLaMA-Factory/.venv/bin/python /home/ll/llm4rec/experiments/run_v34_v43.py --single v38_v29_r3_aggressive_lr7e7_ep10 --gpu <gpu>"
+}