[Bad Case]: 使用LoRA微调后效果很差 #98
Replies: 8 comments
-
我lora微调的效果还可以,下面是参数 deepspeed --include localhost:0,1,2,3,4,5,6,7 finetune.py \
--model_name_or_path /data01/models/open_source/MiniCPM-2B-sft-fp32 \
--output_dir /data01/models/fine_tuned/MiniCPM/cr_6b/ \
--train_data_path /home/MiniCPM/data/cr_6b_glm.json \
--eval_data_path /home/MiniCPM/data/cr_ori_clean_glm3/markdown_2367.json \
--learning_rate 1e-3 --per_device_train_batch_size 2 \
--per_device_eval_batch_size 1 --fp16 --use_lora \
--gradient_accumulation_steps 1 --warmup_steps 100 \
--model_max_length 5000 \
--max_steps 8000 --weight_decay 0.01 \
--evaluation_strategy steps --eval_steps 500 \
--save_strategy steps --save_steps 500 --seed 42 \
--log_level info --logging_strategy steps --logging_steps 10 \
--deepspeed configs/ds_config_zero2.json 我就改了per_device_train_batch_size,增加了model_max_length,修改了deepspeed的配置文件,改成了zero2 |
Beta Was this translation helpful? Give feedback.
-
你用的什么数据集啊,没使用他提供的那个广告数据集吧?我用那个广告的就很差!!! |
Beta Was this translation helpful? Give feedback.
-
我用的自己构造的下游任务数据集,没用广告的 |
Beta Was this translation helpful? Give feedback.
-
抱歉给您带来的不便,之前开源的代码中确实存在 bug,目前已经尝试修复,具体的性能测试还在跟进 |
Beta Was this translation helpful? Give feedback.
-
好的,有消息的话及时通知,我还是挺看好这个模型的 |
Beta Was this translation helpful? Give feedback.
-
你好,能否加一个联系方式,请教你一下微调minicpm,我也是微调下游数据集,比较急,谢谢 |
Beta Was this translation helpful? Give feedback.
-
我没改过微调代码,说实话没啥经验可以交流;我这边的训练环境是8xH800 80G,包版本如下
|
Beta Was this translation helpful? Give feedback.
-
請問dataset的數量是多少? |
Beta Was this translation helpful? Give feedback.
-
Description / 描述
当使用你们提供的LoRA微调脚本和广告数据集进行微调后,输出效果很差(issue区大部分微调效果都不好),我不能确定是否自己写的调用代码有问题,所以请你们提供一个经过LoRA微调后的脚本,或者使用目前的广告数据集微调的话,需要修改那些参数?
Case Explaination / 案例解释
No response
Beta Was this translation helpful? Give feedback.
All reactions