Possibility to fine-tune a model?

[dginovker]

Hello folks!

First, THANK YOU for all your work in the space. Providing an open source, easily runnable model with great UX really lowers the barrier of entry for a lot of people who want to use GPT locally. Every single person who worked on this is amazing 😄

I'm interested in fine-tuning a model. I have a bunch of data that has the generic format of System message\nUser message\nIdeal response. Pretty much what is needed to fine-tune per OpenAI's API docs.

The goal is, because I have this data, the model can be slightly more accurate if given similar prompts to what is in my tuning dataset.

Is it possible to fine-tune a model in any way with gpt4all? If not, does anyone know of a similar open source project where it's possible or easy?

Many thanks!

[cebtenzzre]

manyoso and I are the core developers of this project, and I don't think either of us is an expert at fine-tuning.

My best recommendation is to check out the #finetuning-and-sorcery channel in the KoboldAI Discord - the people there are very knowledgeable about this kind of thing.

All you have to do is train a local model or LoRA based on HF transformers. If your GPU is not powerful then you are probably interested in QLoRA. Once you have a fine-tuned model, you can convert it to GGUF and quantize it using the tools at the llama.cpp repo.