amazing model...can you finetune on a smaller one?
@jondurbin Thanks for this model - follows context really well and is great for RAG use cases. But is too big to run. Would it be possible for you to share a version on mistral v0.2 or something in the 7B parameter range?
Also, can this model be finetuned with cinematika for use in RP, especially in clearly remembering context, following instructions, but with a personality?
Thanks a lot!
Thanks! Yes, I will do another tune of the new mistral 7b base model. This version (and the 7b when it's done) have cinematika in the training data already.
Thanks @jondurbin . Do you have any pointers on how best to use the cinematika dataset or a suggestion on a good dataset to instill a human-like interesting personality, like kunoichi? While most models have a set personality/tone, and some augmented with prompt/finetunes, the fact that airoboros can so strictly adhere to context makes it such an excellent candidate model for a interesting "friend" with different characteristics that evolve over a period of time as part of rapport-building. I plan to inject such character evolution via prompts and specifically like character card, or a daily agenda that it should strictly follow. To populate such an agenda or character card, i was hoping that i could use cinematika or something similar. And to also finetune the model to have a base tone that's always likeable from day 1. Any thoughts on this? Do you have any suggestions on which datasets could help this objective? Thanks again so much!