Official dataset collection of paper: Rethinking Bottleneck in Safety Fine-Tuning of Vision Language Models