1 comment

[ 0.23 ms ] story [ 12.7 ms ] thread
I had to edit the title a bit to fit the character limits, original:

NTK-Aware Scaled RoPE allows LLaMA models to have extended (8k+) context size without any fine-tuning and minimal perplexity degradation.

---------

This is a claimed improvement over the recent Superhot scaling blog post.