You've improved.
I gotta be honest, your first releases gave me the impression you just wanted to churn out a lazy fine tune to give it a funny meme name, and that the result would just be a sloppy less intelligent version of the original.
Well, I tried this one and changed my mind, you finetuned this on Codestral?? I remember when finetunes of CodeLlama-34B were attempted, those were terrible. Congrats, you've improved! This is very good and probably the best at this peculiar size range.
As a minor request/suggestion, please include sampler tips, instruct/context templates, etc so people have a consistent experience.
Well, I tried this one and changed my mind, you finetuned this on Codestral?
... https://huggingface.co/mistralai/Mistral-Small-Instruct-2409
It's not codestral. It's Mistral-Small-Instruct-2409 released a few days ago. But still a goated model.
Codestral really would have been an impressive feat. 🤣
I'm guessing you missed the most recent release from mistral, but there's now a Mistral-Small-Instruct that is a generalist 22B model.
Much easier to work with than the code version (though I'd like to see Drummer attempt to moist-ify Codestral sometime...)
It's not codestral. It's Mistral-Small-Instruct-2409 released a few days ago. But still a goated model.
Man these models get released so fast I didn't even catch that one haha.
Mistral team is absolutely cooking! (And so is TheDrummer). Their 12B, 123B and now 22B models are fantastic, what the hell?
Edit: forgot there's also Pixtral now!
I have to say this model is what I wanted from Theia. Theia was honestly the best out of the nemo tunes that I tried as it followed instructions better than the 12B tunes. This one is great, and I feel like the unslop database was used as well but I have only had one 60 message convo. I'm still trying to figure out the temperature. Also it feels like rep penalty kills this model, and it was repeating a lot of phrases with dry but that was at low temps. At .7 onwards that gets a lot better. I tried temp 5 for SnG and it honestly had normal coherent outputs.
TheDrummer, you've been one of the best tuners out there. I really appreciate everything you do for us plebs.
Yeah, excellent fine tune, especially so quickly after the Mistral 22B release. Last thing missing imo is to tokenize the Metharme keywords (<|user|> and co) for maximum efficiency and it'd be close to perfect.