For short context tasks looks like it's slightly stronger than Llama 7B and slightly weaker than Mistral 7B. Really impressive showing for a completely new arch

For short context tasks looks like it's slightly stronger than Llama 7B and slightly weaker than Mistral 7B. Really impressive showing for a completely new architecture. I've also heard that it was trained on far fewer tokens than Mistral, so likely still room to grow. Overall incredibly impressive work from the team at Together!