Repetition is a common problem with this generation of LLMs, especially the ones that are running on your phone. This is in part due to them being "quantised", which means compressed by removing precision from each neuron.
Sometimes, your characters may say the same few phrases over and over again. To alleviate this problem, you can go to Advanced Settings, and turn on DRY multiplier. The values shown in the picture above a sensible defaults to use. Feel free to tinker with them to get the best results for your character!
For a deeper read on why this happens and how DRY works, see here: https://github.com/oobabooga/text-generation-webui/pull/5677#issue-2177692564
תגובות