-
Currently this repo contains only a single notebook to simply and effectively fine-tune LLaMa3-instruct model with SFT technique on a custom dataset via LoRA.
-
The critical part is data preperation for your own use case. Please note that LLaMa and Mistral family models have different styles of data formatting. In fine-tuned models one may encounter a weird behaviour of generating endless texts! The main rationale is the configuration of EOS tokens and token padding.
-
For experimenting I recommend to try different sets of lora hyperparams and various lora target modules.