nari-router-4b / training_args.json
leapshared's picture
17.16 다섯째 학습: 말을 실기 받아쓰기 꼴로 맞춰 학습(nari_route.stt_form, 서빙도 같은 함수) · 학습 7,141 · 검증 955 · 개발 505, 씨앗 2 · 2에포크 · lm_weight 0.0 · 온도 1.45(합친 판으로 맞춤), 놓친 멈춤 먼저 고름, 2 GB 조각(저장소 routing.md 17.16절)
d2fdf6d verified
Raw History Blame Contribute Delete
509 Bytes
{
"variant": "short_stop",
"lr": 0.0002,
"head_lr": null,
"epochs": 3,
"batch": 8,
"rank": 16,
"alpha": 32.0,
"dropout": 0.05,
"safety_weight": 2.0,
"warmup_ratio": 0.05,
"weight_decay": 0.0,
"grad_clip": 1.0,
"seed": 2,
"grad_checkpointing": true,
"limit": null,
"evals_per_epoch": 2,
"stt": true,
"epoch": 2,
"merged_layers": 248,
"temperature": 1.45,
"lm_weight": 0.0,
"base": "blackmallangcow/nari-decision-4b@029c4628a38859dce5b0db1742e9c97545aabe12",
"temperature_adapter": 1.45
}