georgesung

/

llama2_7b_chat_uncensored

like 367

You take a pre-trained base model like llama-2 (e.g. https://huggingface.co/TheBloke/Llama-2-7B-fp16), then run instruction/conversation fine-tuning on a dataset that does not include alignment and refusals, e.g. https://huggingface.co/datasets/ehartford/wizard_vicuna_70k_unfiltered

To actually run the fine-tuning you can take a look at https://georgesung.github.io/ai/qlora-ift/