Arcee AI's AFM-4.5B

#9
by bartowski - opened

(Disclaimer: I work for Arcee)

AFM-4.5B is particularly interesting because it's a brand new foundational model, using a fresh set of data and some interesting training methods and merging to get good performance, would be nice to see it evaluated on more benchmarks!

https://huggingface.co/arcee-ai/AFM-4.5B

Sure, we will add it to our list. Is this the recommended config for this task?
temperature: 0.5
top_k: 50
top_p: 0.95
repeat_penalty: 1.1

Oh I'm so sorry I somehow didn't notice the notification for this!

Yes that looks reasonable to me :) a lower temperature (like 0.5) seems to work better on this model

We also have support in VLLM now in case that's relevant, as of 0.10.1

Sign up or log in to comment