theblackcat102's picture
Update README.md
c135a1c
metadata
language:
  - en
tags:
  - webgpt
  - regression
  - reward-model
license: apache-2.0
datasets:
  - openai/webgpt_comparisons
  - openai/summarize_from_feedback
metrics:
  - accuracy

Reward Model pretrained on openai/webgpt_comparison and humanfeedback summary. Unlike the other electra-large model this model is trained using rank loss with one more datasets.

On validation dataset the result is much more stable than usual.

You can refer to this wandb for more details

Slightly better than previous webgpt only model : electra-large