Dataset : llama tokenizer 4000 distribution : 80% 13% 7% only for attributes
Chat template
Files info
Base model