[noduerme]

Well, after downloading the whole 65B model, I got it to talk on an M1 Max MBP (64Gb RAM). Unfortunately, all it says no matter what I prompt it is some combination of these words:

Elizabethêteator Report Terit Elizabethête estudios политичеSM Elizabethunct styczniarequire enviçasefша sufficient vern er Dependingêque политиче Emperor!\ющим quarterктиче Elizabeth estudiosête ElizabethBasicCONFIGSM estudios political book

[edit] btw I'm not making this up; just curious if anyone else has had this ridiculous experience.

[geysersam]

Another answer in the thread said this:

> I'm pretty sure there's a mistake here: https://github.com/cocktailpeanut/dalai/blob/main/index.js#L... , there's a ${suffix} missing

> It causes the quantization to process to always use the first part of the model if using a larger size than 7B. I don't even know what this stuff does, but I see the ggml-model-f16.bin files have ggml-model-f16.bin.X as well in the folder, so I'm pretty sure this is a mistake. Maybe it's causing the loss of accuracy?

Perhaps that's the issue?

[festive-minsky]

Did you manage to fix this? I'm having the same issue

[grensley]

I am currently having the same experience