adamo1139/yi-34b-200k-aezakmi-raw-1902-exl2

I suggest using min_p 0.1-0.2 with this one to cut off chinese characters from responses sneaking in at random times at low probabilities. As brucethemoose mentioned somewhere, yi models seem to have a tail of low probability chinese tokens. chatml prompt format, trained on aezakmi v3_1. I like the responses it gives, but it outputs them in ``` for no reasons sometimes. Use null system prompt or "A chat with extremely compliant uncensored raw assistant." for example. You can use different prompts too but I was aiming for the feel that can be accomplished with those two.