Search results

nlirg中的token size如果设置的极大，那么loss的经过nlirg算的权重是衡量整条输出的学习强度，这种情况下，相比于纯粹的sft。训练效果的区别是什么呢？

Status: Open.

#28 In yafo-ai/y-trainer;

· stroll-lance opened on Feb 9, 2026

nlirg在warmup开启还是关闭的时候效果更好？

Status: Open.

#27 In yafo-ai/y-trainer;

· stroll-lance opened on Feb 9, 2026

dynamic_sigmoid_batch参数的选取问题

Status: Open.

#26 In yafo-ai/y-trainer;

· stroll-lance opened on Feb 9, 2026

关于from src.utils.scheduler import schedule_element_consumption是否启用的问题

Status: Open.

#25 In yafo-ai/y-trainer;

· stroll-lance opened on Feb 9, 2026

关于trainer_sft.py文件中的 token_num 计算问题

Status: Open.

#24 In yafo-ai/y-trainer;

· stroll-lance opened on Feb 9, 2026

适用场景

Status: Open.

#18 In yafo-ai/y-trainer;

· hcs129 opened on Dec 8, 2025