How do you decide model settings?
This has been on my mind for a while now, but what's your method of deciding which settings (temp, top_p, top_k, system prompt, etc) to use when testing a given model? My understanding is that each of these can massively impact the output quality and it's not a constant for each model family, so whenever I check out a model from the leaderboard, I'm unsure what kind of setup needs to be done to get similar results to your findings.
I start by checking if a model has official settings and use those and if not use ones from similar models. Sometimes when models have issues with things like repetition (especially thinking models) I have to tweak the settings a bit until things are stable. I also use different settings for reasoning vs writing prompts. I don't list which settings I use for each model because there is no "correct" config, and I would get loads of requests for testing models using various different settings.