exp(dagger-deploy): 轮次20——反馈蒸馏(17.7万样本+3000步微调)负结论:gain=0 时到达仅0.8%(反馈未内化),gain1.8 与基线持平(43.4 vs 43.8)
This commit is contained in:
@@ -0,0 +1,41 @@
|
||||
{
|
||||
"model": {
|
||||
"d": 192,
|
||||
"heads": 8,
|
||||
"ff": 768,
|
||||
"enc": 4,
|
||||
"dec": 4
|
||||
},
|
||||
"train": {
|
||||
"batch": 1024,
|
||||
"lr": 0.0001,
|
||||
"warmup": 1000,
|
||||
"steps": 3000,
|
||||
"wd": 0.01,
|
||||
"betas": [
|
||||
0.9,
|
||||
0.95
|
||||
],
|
||||
"amp": "bf16",
|
||||
"grad_clip": 1.0,
|
||||
"w_mf": 1.0,
|
||||
"w_fk": 0.1,
|
||||
"w_smooth": 0.01,
|
||||
"w_value": 0.0,
|
||||
"w_sub": 0.0,
|
||||
"p_neq": 0.5,
|
||||
"p_edge": 0.1,
|
||||
"p_neq_warm": 0,
|
||||
"fk_tmax": 0.3,
|
||||
"fk_caps": 96,
|
||||
"ema": 0.999,
|
||||
"aug_pre_drop": 0.15,
|
||||
"aug_pre_sigma": 0.1,
|
||||
"aug_dq_sigma": 0.3,
|
||||
"ckpt_min": 10,
|
||||
"eval_every": 1000,
|
||||
"eval_tasks": 256,
|
||||
"log_every": 50,
|
||||
"patience_evals": 3
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user