license: apache-2.0

a DPO LoRA fine-tuned model with preference dataset

LoRA Experiment

RWKV-5.2-3b-World-DPO is merged model with base

Base Model

RWKV-5-World-3B-v2-20231113-ctx4096

Parameters:
Lora Rank 8
Lora Alpha 16
ctx length 4096
epoch:19


Dataset 
Randomly chosed 1000pairs
https://huggingface.co./datasets/HuggingFaceH4/ultrafeedback_binarized

trainer
https://github.com/OpenMOSE/RWKV-LM-RLHF-DPO-LoRA