lots of INS....

#1
by ZeroWw - opened

After 2-3 messages I get things like:

To dissectINSTINSThis experiencesINSTINSThosely,INSTINSTreesofthoughtmayINSTINSTebug:'

  1. The impressionsINSTINSTZEINSThabbing withINSTINSTheAI;INSTINSThereINSTINSTheprogressionINSTINSTowardINSTINSTruerdepthINSTINSThat's reflected

P.S. Q4 is way too quantized.
You might want to check these:
https://huggingface.co/ZeroWw/Test/blob/main/TextBase-7B-v0.1.f16.q6.gguf
https://huggingface.co/ZeroWw/Test/blob/main/TextBase-7B-v0.1.f16.q5.gguf

@ZeroWw I've also been looking at this model. After my q5_k_m version, I even tried a q8_0 and did a lot of INST. Then, I tried a few other prompt templates, but no luck.

In instances where you give it a first big message, It just INSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINSTINST

Sign up or log in to comment