There are some issues with fine-tuning here

[Lufffya]

1, Unable to reproduce the ability to write code using official Lora

2, There were some issues with the model and data fine-tuning using this command line

python finetune.py \
    --base_model 'decapoda-research/llama-7b-hf' \
    --data_path 'yahma/alpaca-cleaned' \
    --output_dir './lora-alpaca'

Output carry instruction mark:

Repeated input instructions：

Output some incomprehensible content:

I guess the model training may have broken down Whether I should change some hyperparameter when fine-tuning, instead of fine-tuning in the simplest way by default, but I don't know where to start

[lywinged]

Check my test.py in PR, it can explain all these cases.

[Lufffya]

Check my test.py in PR, it can explain all these cases.

Okay, thank you very much. May I ask how you resolved this situation

[lywinged]

Check my test.py in PR, it can explain all these cases.

Okay, thank you very much. May I ask how you resolved this situation

I read all the issues both closed and open, and the solution are listed there.

[Lufffya]

@lywinged Okay, I observed the training and testing examples by running python test.py, But I still don't understand what's going on. Sorry

[Oxi84]

Maybe because model decapoda-research/llama-7b-hf has bad tokenizer. Try the one from finetune doc, there were some issues with llama tokenizer reported on huggingface

[kuan-cresta]

I had this issue earlier and figured it was due to some issue saving the final model: adapter_model.bin

ls -l
-rw-r--r-- 1 root root      443 Apr 21 19:29 adapter_model.bin

Note the file size is only 443.

I manually copy the checkpoint pytorch_model.bin and use that.

ls -l
-rw-r--r-- 1 root root 67201357 Apr 26 18:36 adapter_model.bin

[WoodScene]

@kuan-cresta , your method is correct! Thank!

[paolorechia]

Thanks for sharing the solution, @kuan-cresta! Copying the checkpoint binary solved a similar problem for me too

[su-park]

Even though I replaced adapter_model.bin with a checkpoint binary as @kuan-cresta mentioned, there have been some improvements, but the same issues persist. Do you have any more suggestions?