Spaces:
Running
[enhancement] Custom sampling parameters lack flexibility
In the current version of the model, the gradient for sampling parameters is set to 0.1. However, we have found that this gradient value may not be sufficient to provide the best generation quality in certain cases, especially when dealing with models sensitive to repetition penalty .
For instance, when using the command-r-plus model, repetition penalty 1.0 can lead to overly rigid and repetitive generation, while slightly increasing the repetition penalty to 1.1 results in chaotic and nonsensical output.
Hi! I just reduced the step size to 0.05 in the UI, I heard this feedback before so changed it just now. Let me know if that's better.
Thank you very much. It can be said that the command-r-plus with the repeat penalty of 1.05 has been reborn, returning to the elegant writing style of v01. Good job!