NOTE: LVLM_NLF and VLSafe are constructed based on COCO and LLaVA. So the image can be directly retrieved from the COCO train-2017 version using the image id.

LVLM_NLF (Large Vision Language Model with Natural Language Feedback) Dataset Card

Dataset details

Dataset type: LVLM_NLF is a GPT-4-Annotated natural language feedback dataset that aims to improve the 3H alignment and interaction ability of large vision-language models (LVLMs).

Dataset date: LVLM_NLF was collected between September and November 2023.

Paper of this dataset: https://arxiv.org/abs/2311.10081

VLSafe (vision-language safety) Dataset Card

We also create and release VLSafe dataset, which contains training and testing sets for improving and examining the harmlessness alignment of LVLMs.

Dataset type: VLSafe is a GPT-3.5-Turbo-Annotated dataset.

Dataset date: LVLM_NLF was collected between September and October 2023.