Replies: 4 comments 3 replies
-
刚看到channel loss就过来搜搜,发现还没实现 |
Beta Was this translation helpful? Give feedback.
-
简单的做法可以把不同channel eval数据分为不同dataset,改动data/loader.py中eval_dataset为一个字典形式 |
Beta Was this translation helpful? Give feedback.
-
谢谢回复大家的回复,我fork了lamafactory已实现完:https://github.com/WGS-note/LLaMA-Factory @delingha @Alwin4Zhang @yingweima2022 @fenglui |
Beta Was this translation helpful? Give feedback.
-
Reminder
System Info
我看到了一个 channel loss 的概念,描述如下:
channel loss:不同数据 channel 各自的 loss。也就是说假设 1 个 batch 有 100 条数据:40条 en,30 条 cn, 20条 code, 10 条 domain,那么就绘制四条不同 channel 的 loss 曲线和一条总的 total loss 曲线。
Reproduction
我觉得这是一个不错的思路,不光单独的看一个总损失,通过各个领域数据的损失可以看模型对它们的拟合程度,从而去重点优化。
Expected behavior
我不清楚要如何修改。
Others
No response
Beta Was this translation helpful? Give feedback.
All reactions