This is the dataset used to do the initial training for JoyCaption Beta One (https://huggingface.co/fancyfeast/llama-joycaption-beta-one-hf-llava), before post-training.

Contents

Most of the dataset focusses on descriptions and captions for images, with a smaller subset covering general VQA tasks.

Some of the questions and answers are human written, some are automated, some are machine written. The is_human column is True when the answer text is human written.

WARNING

This dataset is meant for research purposes. It is generally unfiltered, and contains content submitted by other people. I cannot guarantee that is free of offensive materials.