Silkie is a visual language model trained using preference distillation on GPT-4V annotated AI feedback. It is a fine-tuned version of Qwen/Qwen-VL-Chat and was trained on our MMInstruction/VLFeedback dataset with direct preference optimization (DPO). Silkie is a visual language model trained by preference distillation on GPT-4V annotated AI feedback. It is a fine-tuned version of Qwen/Qwen-VL-Chat that is trained on our MMInstruction/VLFeedback dataset with direct preference optimization (DPO). Compared with the original model, Silkile achieves 6.9% and 9.5% relative improvement on the MME benchmark regarding the perception and cognition capabilities, respectively. Besides, Silkie sets a new state-of-the-art score of 3.02 on MMHal-Bench regarding hallucination evaluation. Please refer to our project page for more details.

Silkie is intended for research purposes, particularly for alignment research in multimodal models.

Below is a simple Python code snippet to get started with the model. For installation instructions please refer to our Github repository.

@article{2023vlfeedback,
  author      = {Lei Li and Zhihui Xie and Mukai Li and Shunian Chen and Peiyi Wang and Liang Chen and  Yazheng Yang and  Benyou Wang and  Lingpeng Kong},
  title       = {Silkie: Preference Distillation for Large Visual Language Models},
  publisher   = {arXiv:2312.10665},
  year        = {2023}
}