README.md

metadata

license: apache-2.0
datasets:
  - opencsg/UltraFeedback-chinese
language:
  - zh
base_model:
  - opencsg/csg-wukong-smoltalk-chinese

• We apply DPO to opencsg/csg-wukong-2b-smoltalk-chinese, using ultrafeedback-chinese-binarized-lowest as the DPO dataset.
• Train 2 epochs, learning rate=5e-6, batchsize=128
• This model's evaluation score is higher than that one trained on ultrafeedback-chinese-binarized