Feedback
#2
by
ThiloteE
- opened
I tried both the Q4_0 and Q8_0 quants in lmstudio (version 0.2.18) and liked this model :-) Great job!
What I found was this:
- EOS Token issues. It responds with
</s>a lot, even when I tried to filter them out with LMStudio. This might be be caused by quantization.
- A temperature of "0" gives awful responses, as it seems to spam "#######" (At least I think it was the model. Hope it wasn't a bug in LMStudio). Anyway, better increase it to higher ones like 0.9 or something.
The model is obviously not perfect, as I have seen some instances where it did not rephrase a certain name, but it is not all too bad :-)
Is there a chance you will do another finetune of this on one of the more recently published models?
Are there plans to release the dataset at one point? That would be a great service to the open source community and who knows, maybe somebody will pick up the torch and improve upon it?