Note that this is StableLM ALPHA (only 0.52 epochs into training).

The fully trained version will surely be much better.

Also, you should benchmark GPT-3 Babbage for a fair comparison since that is the same size as 7B.