RLHF-Reward-Modeling

• This is a fork of RLHF-Reward-Modeling
  • support models which can handle Japanese
  • support Unsloth, which reduce VRAM when training and accelerte training efficiency
  • support wandb

Support

model	support
google/gemma-2b-it	✅
llm-jp/llm-jp-3-1.8b-instruct	✅

dataset	support
hendrydong/preference_700K	✅
xxxx	-

Environment setup

git clone https://github.com/ohashi3399/RLHF-Reward-Modeling.git && cd RLHF-Reward-Modeling

Bradley-Terry-RM

export HUGGINGFACE_API_KEY=<Your HUGGINGFACE_API token>
export WANDB_API_KEY=<Your WANDB_API token>
source setup.sh && cd bradley-terry-rm
source tune_bt_rm.sh