Closed 0x000011b closed 1 year ago
Yep, about to test it out. I haven't seen setting top-k to 40 anywhere, though. Does that personally give you better results, or did you see it from somewhere?
It's a leftover way back from when I was originally testing by manually doing inference with some copy-pasted example HF code, but it seemed to give me better results when testing on the 6B. Feel free to play around and suggest other defaults.
Now that we've gotten some feedback and played around a bit with the models, I think these are some better defaults to bundle on the UI. @TearGosling Can you test them and let me know what you think?