3 ms·
Yes, but post training cannot possibly account for all possible use cases. Sane defaults are fine, you can't really do much about sampling parameters in chatbot
by curioussquirrel 6mo ago
Yes, but post training cannot possibly account for all possible use cases. Sane defaults are fine, you can't really do much about sampling parameters in chatbots and coding harnesses anyway. And when making an API call, you have to actively change the parameter in your payload. I don't believe there's any real risk.
- charcircuit 6mo agoThe risk is that people tweak it, potentially on accident, and then think the model is bad instead of understand they are using it wrong. This causes potential reputational damage by exposing the control.