search

ianarawjo / ChainForge

An open-source visual programming environment for battle-testing prompts to LLMs.
https://chainforge.ai/docs
MIT License
2.2k stars 169 forks source link

Support for fine-tuned OpenAI models #136

Open ekiwi111 opened 11 months ago

ekiwi111 commented 11 months ago

It's currently non-trivial to add a new model to the "Model Version" dropdown in the model settings.

Adding a new value to the enum property in ChatGPTSettings, followed by adding a new entry to the NativeLLM enum and then rebuilding the frontend, is not sufficient. I'd like to add a new gpt-3.5-turbo-instruct model and a custom fine-tuned one.

What are the steps?

ianarawjo commented 11 months ago

There’s already a feature for this on our docs called custom providers: https://chainforge.ai/docs/custom_providers/

ianarawjo commented 11 months ago

Per the instruct point though, I do think I need to add that in a new update. OpenAI changes things so fast can’t keep up to date —that’s why custom providers exists, for edge cases like these. I’m currently resting on a semi-vacation after a tumultuous deadline, but will return next week.

coolk8 commented 10 months ago

Hi! Thank you for a great tool.

Just want to report that new OpenAI models are not working right now in https://chainforge.ai/play/. I can choose them but got an error on run:

gpt-4-1106-preview: "Language model gpt-4-1106-preview is not supported. gpt-3.5-turbo-1106: "Language model gpt-3.5-turbo-1106 is not supported."

ianarawjo commented 10 months ago

Thank you for alerting me to this. Fixing now.

ianarawjo commented 10 months ago

The problem is fixed. New versions are on /play and pushed to PyPI.

ianarawjo commented 10 months ago

I am keeping this open to prioritize adding support for fine-tuned OpenAI model endpoints, natively in the settings window. If anyone wants to submit a PR for this, too, it would be appreciated.