HiTZ/rmloss-opt-rm-13b

LM Loss OPT RM

This is a fine tuned OPT 13b model for reward modelling. The finetuning has been done on top of the full SLF5K dataset following the method presented in the paper Training Language Models with Language Feedback at Scale. The main results can be seen in the following table:

Model	# Params	Validation Accuracy (in %)
OPT LM Loss	13B	73.4 +/- 1.9
OPT LM Loss	1.3B	69.6 +/- 2.0
OPT RM Loss	13B	71.8 +/- 2.0

If using this model, please cite the following paper:

@article{scheurer2023training,
  title={Training Language Models with Language Feedback at Scale},
  author={Scheurer, J{\'e}r{\'e}my and Campos, Jon Ander and Korbak, Tomasz and Chan, Jun Shern and Chen, Angelica and Cho, Kyunghyun and Perez, Ethan},
  journal={arXiv preprint arXiv:2303.16755},
  year={2023}
}