Chinese token capabilities?
#1
by
at676
- opened
Hi, I saw that you quantized Yi 34B with QuIP#, exciting to see our stuff being used. I was playing around with this model on interactive_gen.py and it doesn't seem to be very good at handling chinese tokens. Did you generate hessians with the default redpajama dataset we hardcoded in hessian_offline_llama.py? If so, you'll want to generate hessians with a chinese + english dataset to get accurate hessians for quantization.
Interesting, maybe my terminal (mintty msys2) is not sending chinese characters correctly.