GGUF
Not-For-All-Audiences
nsfw
Inference Endpoints

MistRP-AirOrca

#1
by Huegli - opened

Hey Undi95,

just a short feedback: This model feels surprisingly pleasant. Ok, it isn't as 'intelligent' as most of your 13B models are, or 20Bs. But for a 7B model (q8) it performed pretty impressive in my first easy test, and it isn't stubborn to take corrections. It tends to be a bit repetitive in the story line, but this might be my fault, I have to test more. Or let me put it the other way round: I've seen 13B models performing worse than this one.

I have a request: You put out new models in such a high pace that it becomes hard to keep track of what the underlying models capacities are. Would you please add a suggestion for the maximum usable context size to the model cards? As long as you used llama-2 as base model, we could easily use 8k as a very reliable context size. But with your newest merges I have lost track of what the models are capable of in terms of context size.

As always: Thank you for your great effort. I can't even imagine how you do all this that fast.

Greetings,
Huegli

Sign up or log in to comment