That is awesome!
#4
by
owao
- opened
You did it! Given the power of diffusion models on so many domains I can't wait to see what we can do around text generation.
I was keeping an eye on this to happen and this really is the first one I see surpassing GPT-2 level!! And I mean even closing the gap with llama3-8b, that's been crazy fast! Congrats to the team!
For real, I didn't expect diffusion models to become even halfway decent, and we got LLaMA 8B level performance with a license which is actually permissive (something you can't expect from LLaMA itself, and neither from other models with its performance).
Huge props to the team =)