Simple Q&A Fine Tune Dataset

#40
by faheemraza1 - opened

The script sample_finetune.py uses HuggingFaceH4/ultrachat_200k dataset. This dataset has turn by turn conversation data in it. How important it is for LoRA fine tuning this model to have a turn-by-turn dataset? I have some books related to a very niche engineering field. It would be hard to generate a turn-by-turn dataset from them. Can I fine tune using a simple Q&A styled dataset? How effective will that be? And is there any chance that it will kill the model's ability to make back and forth conversation?

Sign up or log in to comment