Good!
Had a little play with it. Prose seems similar to the two other models capybara and zephyr (which is to say, more compelling than the vast majority of Mistral models and probably as good as any high rated 13b) - perhaps a little more down and dirty, and perhaps a little more aware of context and how things work.
It seems slightly smarter though, like it's pulling info from the instructions and character card, which I haven't really seen Mistral do properly before, even the models with lima rp or pippa - not even from the mistral merges and tunes that are supposed to be good at instruct. Which is honestly the area where 7b's are the weakest (instruct and comprehension), so a good win probably. Usually when I see any mistral or 7b merge referencing my cards or instructions it's literally misinterpreting it somehow (and that's pretty rare). Here it did it a quite few times, correctly. ofc, it's not doing this consistently but it's 7b what do you expect lol!
Wasn't sure if I'd ever see a 7b model really do that even some of the time properly.
Overall, really good actually on first test. If you find time to put up the HF python model I'd totally download that in case there are any future merges that might be worth playing with.
Yeah, solid I think. Feels like the 7b version of Stheno or Mlewd-xwin tbh. Obviously can never be quite as good as a 13b, but boy does it feel close. Finally feels like that stage we got to with 13b, where it's a real high quality merge that has all the right elements.
Also, and I'm not sure if this was dumb luck, but I seemed to get less repetition than other mistrals. Certainly no identical replies which for Mistral is common for pretty much anything I've tried.
Edit after a month: The prose IS really good IMO, combination of the base models and loras. However I might have overstated any cleverness. It does sometimes pick up from cards, which admitted most other models do not do well, but probably it's instruct following is still an area for improvement.
Good then! Haha! It was done for you in the first place, hope you enjoy it
For a 7B it is very creative and really less repetitive than 7b models. Is there a recommended prompt format or settings to stop it from writing on behalf of the user?
For a 7B it is very creative and really less repetitive than 7b models. Is there a recommended prompt format or settings to stop it from writing on behalf of the user?
You can try ChatLM prompting, or fastchat, but not sure. I use Alpaca, it work fine
Good then! Haha! It was done for you in the first place, hope you enjoy it
I certainly am! It's good shit. Much appreciated.
Maybe there's an 11b in here somewhere too one distant future day ;) (I can just barely run those)
Interestingly, this merge shot to the top of Ayumi's leaderboard, even among 70Bs. Its #3 on IQ alone:
http://ayumi.m8geil.de/ayumi_bench_v3_results.html
Maybe that task arithmetic paper is onto something.
Interestingly, this merge shot to the top of Ayumi's leaderboard, even among 70Bs. Its #3 on IQ alone:
http://ayumi.m8geil.de/ayumi_bench_v3_results.html
Maybe that task arithmetic paper is onto something.
Lmao, nice
i will be using this after finding out the xwin mlewd merge didnt go as planned, i only just found that post earlier today. ill try to report back with qualitative comments
i found this model by chance. This merge is truly a great work. I am using it in silly tavern for storys, and it does well. Actually it is my favorite model. I am using the q8 variant.
my requirements for models are:
- coherence
- simulation of the characters' way of speaking
- anatomy knowledge for violence like fantasy fights and their effects
- some porn skills for all cases. you never know :)
- simulate characters based on character sheets and situations.
- description of environment, characters etc.
Toppy does it all well for me. I also used the other models from the top of the charts like the go bruins merges.
My second favorite model is quantumaikr_quantum-dpo-v0.1 but Toppy is clearly above it. This Toppy merge is a gold nugget in the river. thanks for it.
Once again, this is the pinnacle of nsfw 7b models. Bravo Undi.
Once again, this is the pinnacle of nsfw 7b models. Bravo Undi.
Thanks a lot!