Fine-tuning Minitron-4B-Base?
#2
by
vaclavkosar
- opened
Recently, Nvidia published this 4b distrilled model with some evals equal to the 8b.
https://huggingface.co/nvidia/Llama-3.1-Minitron-4B-Width-Base
If you guys fine-tune that to a Dolphin chat model, it may become one of the first fine-tune of this kind of performance level within 4b size category.
There is this model also: https://huggingface.co/solidrust/Llama-3.1-Minitron-4B-Magpie-SFT-800K-MT-Magpo-3.1-Pro-05-AWQ
Only concern if there are some architecture changes. I think I saw a conversion of Nemo to Llama architecture for compatibility somewhere.