original model gone / output bad for 16k context?

#1
by mclassHF2023 - opened

It seems the original unquantized model is gone?
Also, with all wizardLM-2 7B models, I notice that at (maybe before) 16k context, the output is just garbage, things like:
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1
or even simply newlines, no real output at all.

mclassHF2023 changed discussion title from original model gone? to original model gone / output bad for 16k context?

It seems it is gone. As is the original wizardlm-2.

At least with smaller context it works fine though so it's not broken.

Sign up or log in to comment