How to finetune little llama base model with dataset or rawtext

[lbarasc]

Hi,
I want to finetune a little llama base model with custom data (in french language). I have Xeon e5, 64 Go and RTX 3060 12 Go.
Which mini llama model (or other multi lingual small model) can i use ?
What type of custom dataset should i use ? where can i find a sample dataset with the right format ? what is the right file format (.json, .jsonl ?) file seems to not appear in oobabooga (i copy the file under training\datasets)
i try a lot of things but i have error when i try to finetune with custom dataset (oobabooga error is like "error only training llama, gpt-x...)
Finetuning with raw text seems to work for me.
Thank you for your help.