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.
Base model: https://huggingface.co/OpenRLHF/Llama-3-8b-sft-mixture
Cosine Scheduler
Learning Rate: 9e-6
Warmup Ratio: 0.03
Batch Size: 256
Epoch: 1