Simuletic/Surveillance-VLM-Weapon-Knife-Detection-Dataset

这是一个名为Simuletic Safety VLM Dataset的开源数据集，专为视觉语言模型（VLMs）的指令调优设计。数据集旨在教导模型定位武器（刀具和枪支）、推理威胁并避免误报。包含高质量的样本，涵盖三个关键类别：刀具检测、枪支检测和安全/阴性样本。数据集采用标准化的0-1000坐标系统，与Qwen-VL/LLaVA兼容，并包含链式思维安全推理。数据以行业标准的JSON对话格式存储，适用于多模态LLMs（如Qwen-VL、LLaVA、PaliGemma）的微调、安全性和基础能力的基准测试，以及自主监控代理的研究。数据为计算机生成，无真实人类记录或伤害，采用CC BY 4.0许可。

This is an open-source dataset named Simuletic Safety VLM Dataset, designed for instruction tuning of Vision Language Models (VLMs). It teaches models to locate weapons (knives and firearms), reason about threats, and avoid false positives. The dataset includes high-quality samples covering three critical categories: knife detection, firearm detection, and safe/negative samples. It uses normalized coordinates (0-1000 scale) compatible with Qwen-VL/LLaVA and incorporates chain-of-thought safety reasoning. The data is pre-formatted in a standard JSON conversation format and intended for fine-tuning multimodal LLMs (e.g., Qwen-VL, LLaVA, PaliGemma), benchmarking safety and grounding capabilities, and researching autonomous surveillance agents. The data is computer-generated, with no real humans recorded or harmed, and is licensed under CC BY 4.0.