Open ucalyptus2 opened 5 months ago
cc: @jon-tow @hardmaru @estability
I am curious about that part very much - tinkering with StableLM and trying to find a way to teach the assistant its' own name.
P.S.> a dirty hack that worked for me was running LORA for 3 epoch with just 16 records in the dataset (on top of SFT instruction tuned base model) that had the prompts with different versions of "what is your name" conversation. It worked yet did hit the overall quality of responses.