通过LLaMA-Factory代码预训练,Loss不降反增,训练之后回答混乱 #612
-
训练环境:Windows 10 训练指令(通过LLaMA Factory微调):set CUDA_VISIBLE_DEVICES=0 数据集:StarCode的单个C代码文件,qarquet格式 训练日志:训练后效果:我的问题是:
|
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
不支持code 能力微调 |
Beta Was this translation helpful? Give feedback.
不支持code 能力微调
loss不降大概率是数据构建有问题 llama factory我们测过了是正常的