GPT-4o usage - Githubissues

ubiquibot / conversation-rewards

0 stars 10 forks source link

GPT-4o usage #39

Closed Keyrxng closed 3 weeks ago

Keyrxng commented 1 month ago

Resolves https://github.com/ubiquibot/conversation-rewards/issues/30

removed js-tiktoken due to the 128k token limit and strict GPT-4o usage
making one call instead of a batch of 10

0x4007 commented 1 month ago

Yeah the primary work here is to simplify the code.

No more choosing an "optimal" model and no more sampling/averaging scoring!

Only use chatgpt 4o as it is superior in every way.

We also always strive to simplify the codebase whenever possible. Thanks @Keyrxng

0x4007 commented 3 weeks ago

@Keyrxng I assume you tested that this works.

Keyrxng commented 3 weeks ago

@Keyrxng I assume you tested that this works.

Truthfully no I did not QA it on a test issue, I didn't think it was necessary because this is for the V2 bot which is doing test scoring on each closed issue anyway, my apologies.

0x4007 commented 3 weeks ago

Guess we will see if it works @gentlementlegen be sure to include these changes so we can test asap.

gentlementlegen commented 3 weeks ago

@0x4007 Should be using it from now on. I noticed that because we set development as the default branch, when not specifying the branch on the .ubiquibot-configuration.yml it uses the development branch for runs which might not be wanted (I think we should rely on the main branch for production). https://github.com/ubiquibot/conversation-rewards/actions/workflows/compute.yml

0x4007 commented 3 weeks ago

It should always check the default branch and use that as the default. Given that we are rapidly iterating now, it makes sense for development to be our default branch. Later in the future when things are stable we can consider switching to main as the default.