Finetuning Code

[kaifronsdal]

In your paper, you mention fine-tuning models on the data generated in this repo. Would it be possible to make the model fine-tuning and evaluation code avaiable?

[bammt]

Thanks for your question. However, we think we don't need to add the fine-tuneing code because we just use BELLE: https://github.com/LianjiaTech/BELLE to fine-tuning our llama models. As for evaluation, we had released them. Just use test_correction_direct_vllm.py for all-direct method reasoning. test_correction_sub_step_check_vllm.py for step-direct and step-cot method reasoning.

[kaifronsdal]

I could be mistaken but it seems like the evaluation code seems to be setup for the base llama2 models, not fine-tuned models?

[bammt]

I have updata my code, you need to replace the 'check_path' with your own fine-tuned ckpt.

[kaifronsdal]

Ah, thank you so much!

[bammt]

You're welcome. And if I have solved your problem, could you please give me a star? thx : )