Skip to content

Latest commit

 

History

History

data

README.md

The source of the dataset is described below:

  • Image Folder: Download 2017 Train images from https://cocodataset.org/#download.

  • SFT Dataset: Download from LLaVA-Instruct. Like LLaVA, we can also pre-train for feature alignment using SFT datasets, such as CC3M.

  • RM Dataset: Download from LLaVA-Human-Preference. We use the convert_to_llava_reward_dataset.py script to convert it into the format of our reward training dataset.

  • PPO Dataset: We use the same data format as SFT, but in this answer field we can leave it empty or give it a random string.