v2 - thoughts
Continued from https://huggingface.co/jukofyork/command-r-35b-writer/discussions/1
Seems to work overall. I've noticed small issues with it.
- getting clothing wrong in descriptions
- orientation of characters
- some minor misspellings
got a Japanese word once but with temp under 0.8 i haven't seen a repeat.
With multiple characters talking it can easily output hundreds of tokens in text that looks... pretty darn decent.
Thanks for the feedback! I'm about 2 days off the new qwq-32b-v2 finetune and then will likely try again for the v3 version of this model.
Mhmm... Going a bit longer i've gotten a few Japanese words thrown in, usually referring to body parts.
Not sure if you've seen it but I uploaded v3 here a couple of days ago:
https://huggingface.co/jukofyork/command-r-35b-writer-v3
It's mostly fixed the weird spacing after newlines, but now seems to add weird extra spaces after period symbols instead :/
In theory it should be much better than the v1 or v2 models as it was trained on nearly 2.5B tokens (around 10x more).
Not sure if you've seen it but I uploaded
v3here a couple of days ago:
Yeah i see an mradermacher quant. I'll get it and try it soon. :)