chuyi777's picture
Update README.md
595c2d5 verified
|
raw
history blame
277 Bytes

The Llama3-8b-based Reward Model was trained using OpenRLHF and a combination of datasets available at https://huggingface.co/datasets/OpenLLMAI/preference_dataset_mixture2_and_safe_pku.

Cosine Scheduler
Learning Rate: 9e-6
Warmup Ratio: 0.03
Batch Size: 256
Epoch: 1