{ "created_at": "2026-07-27T23:14:42+08:00", "name": "v102_v74_pcgrad_normalized_task_delta", "method": "offline_parameter_delta_pcgrad", "method_caveat": "PCGrad is applied to specialist parameter deltas, not simultaneous minibatch gradients. This preserves the conflict-projection rule while making the experiment reproducible from existing experts.", "anchor": "/home/ll/llm4rec/experiments/outputs/v74_public091_rslora_loraplus_r128_a16_lr1e5_x16_ep3", "experts": { "user": "/home/ll/llm4rec/experiments/outputs/v84_public091_itemic_w2_rslora_loraplus_ep3", "video_ad": "/home/ll/llm4rec/experiments/outputs/v87_v74_final_itemic_focal_repair_r32_lr1e6_ep025", "product_live_world": "/home/ll/llm4rec/experiments/outputs/v92_public091_itemic_io_lora_plus_body_r128_ep3", "ad_world": "/home/ll/llm4rec/experiments/outputs/v56_public091_rslora_r128_a16_lr6e5_ep2" }, "formula": "For layers 8-27 projection matrices: top-10% specialist deltas, normalize each tensor to the v87 delta norm, apply deterministic pairwise PCGrad, average at 0.25, cap at 0.30*v87 norm.", "changed_tensors": 140, "injected_global_l2": 0.005041098032500604, "sources_sha256": { "v74": "e626e97abf4a373809fa30f5909bbc44db5290cfb5d6576eb1ab37ce7165d6bc", "user": "98115e06e232cf1c014985b637c8411dcdcede35edb709b0d902e7ac3e769eaf", "video_ad": "3b310613c97d2d8af1b8f33fa7b5b8dc9b0da6b251f74355f1ac0a7fa378a5db", "product_live_world": "53ae3ba523780ed5f35cc72f7cc462a481480037266ea5dbe3ae6d719537f523", "ad_world": "ca22045c0460b7914fe1aefdda9e4f74921bd61ae5a664e1db53bd0ed3dd4bfa" }, "checkpoint_policy": "Derived final model only; no intermediate checkpoints.", "script": "/home/ll/llm4rec/experiments/build_v102_pcgrad_task_delta.py", "script_sha256": "39af8c57d0a712fc0f5dad29aa130cd8985d86328e39dc425a11e151fdb6b756", "reproduce": "demo/LLaMA-Factory/.venv/bin/python experiments/build_v102_pcgrad_task_delta.py", "output_sha256": "69c70cdf66bf61cfb32890e88fe56116da952e4315c0059b019b724e61f7e7c1" }