Feature request

The TRL library introduced support for NEFTune in this pr

Motivation

Improves Supervised Fine-tuning performance

See paper : https://arxiv.org/abs/2310.05914

Your contribution

I can try to port code over from TRL.