zephyr_0.1

The DPO-trained model from alignment-handbook/zephyr-7b-sft-full using 10% data of HuggingFaceH4/ultrafeedback_binarized, as in the "Weak-to-Strong Extrapolation Expedites Alignment" paper.