Coding Evaluation

[kygguo]

Thanks for sharing this awesome repo!

When I try to reproduce the results of Eurus-7b NCA in paper, I find that compared to Eurus-7b SFT

1. the coding scores decrease consistently for all three tasks.
2. all pot scores are much lower than the corresponding cot ones.

I used the lastest code in this repo, could you provide any suggestion on this?