Finetune on unstructured dataset

[lun-4]

I noticed that the finetuning scripts assume that the dataset is going to be instructional, but what I plan to do isn't such, so I took it to draft out an implementation of dataset preparation and relevant changes to LLaMA-Adapter to support such tuning.

The dataset preparation was copied from prepare_alpaca.py, which then had a lot of its setup stripped out of the instructional details.

I'm testing this code at the moment by training a model, it seems to be working well, but I'm opening this code to discussion and review beforehand.

[lantiga]

Thank you @lun-4, looks good!

Could you:

• add the same instruction_tuning parameter (defaulting to True) to lora.py and full.py?
• also add the corresponding CLI option to the finetune_ scripts?

Thanks a lot!

[lantiga]

Thank you @lun-4, merging! A quick howto would be super appreciated :-)