Randomly switches to Russian and even inserts gibberish english on Q6_K on KoboldCPP + Silly Tavern. Mistral Instruct prompt format.
EDIT: This is still happening even on the Q8_0 quant.
I am asking it simple questions as an assistant. For example, I ask it who is the best mixed martial artist of all time and it starts decently well, but somewhere along the way, it starts answering in Russian and then turns into complete gibberish.
Also the times it does respond appropriately ( to other questions), it is extremely ( and unnecessarily ) verbose. It beats around the bush a lot with unnecessarily complex sounding vocabulary, but it's like its being vague just for the sake of it.
I am using the Mistral Instruct prompt format.
Ah, I noticed it once with an extremely specific instruction. It might be some weird trigger word that switches it to Russian/gibberish. Would you mind sharing your instruction?
I had asked it 'Who is the best mixed martial artist?' and I think it started switching to Russian after mentioning Khabib Nurmagomedov ( who is Russian )
Is it something that happened once or does it happen every time you ask this question? I couldn't reproduce it in 10+ attempts.
This kind of gibberish / Russian shows up on my end often.
Yet again
Ok cool it looks like it can be solved by tweaking the inference parameters. Here's what LM Studio uses:
@siddhesh22 i saw this in some models when the min p is too low , just increase the min p set it to 0.05 or 0.1~ it will be solved .
Ok cool it looks like it can be solved by tweaking the inference parameters. Here's what LM Studio uses:
Thank you! It's working really well now. Sillytavern has a 'neutralize samplers' setting which I was using previously which was causing the issue ( but it worked fine with other models) and after adjusting the values, I am getting much better results, the model is amazing!
Thanks, glad it solved this problem! :)