21 lines
1.0 KiB
JSON
21 lines
1.0 KiB
JSON
{
|
|
"created_at": "2026-07-13 03:38:30",
|
|
"name": "v33_task_arith_v29_live55_r325",
|
|
"method": "task_arithmetic",
|
|
"formula": "theta_v29 + 0.55*(theta_v31-theta_v29) + 0.25*(theta_v32-theta_v29)",
|
|
"base": "/home/ll/llm4rec/experiments/outputs/v29_v19_user_world_guard_lr8e7_ep018",
|
|
"experts": [
|
|
{
|
|
"path": "/home/ll/llm4rec/experiments/outputs/v31_v29_live_specialist_r3_lr12e6_ep035",
|
|
"task_vector_scale": 0.55
|
|
},
|
|
{
|
|
"path": "/home/ll/llm4rec/experiments/outputs/v32_v29_balanced_r3_draft_lr8e7_ep020",
|
|
"task_vector_scale": 0.25
|
|
}
|
|
],
|
|
"rationale": "Approximate OneReason specialize-then-unify without leaderboard labels: retain 20% explicit v29 weight, emphasize the live-domain expert, and add a smaller balanced R3 cognition delta.",
|
|
"cot_policy": "All checkpoints descend from the CoT-native v29/v19 line; v7 is not used.",
|
|
"reproduce": "demo/LLaMA-Factory/.venv/bin/python experiments/task_arithmetic.py --name v33_task_arith_v29_live55_r325 --overwrite",
|
|
"runtime_seconds": 2.11
|
|
} |