More flexibility in parameters for OpenAI / LiteLLM

[satpalsr]

1. Why not expose entire LiteLLM config as parameter or something when running evals?
2. I'm actually looking to run open-r1 evals with any api provider. This don't seem possible with current. The OpenAIClient class should also allow using any base_url. Parameter like temperature should be available to modify directly.

[gary149]

+1 also it would be great to have examples of usage with openai endpoints and LiteLLM.

[julien-c]

and Inference Providers on the Hub

[lhl]

I saw @NathanHB and @gary149 just committed some improvements so that various configuration options can be submitted via a YAML file (gj!), which is workable, but I think that proper flags or a way to submit via command line would still be useful for building eval pipelines (vs having to add YAML file generation into my scripts, which is a big bloat/PITA).