Nemotron-3.5-Content-Safety-Dataset
收藏资源简介:
Nemotron 3.5内容安全数据集是一个用于人机交互内容安全分类的混合真实/合成监督指令数据集。该数据集由英伟达公司于2026年6月创建,遵循CC-BY-4.0许可证,可用于商业用途。数据集包含纯文本和基于图像的单轮对话示例,每个示例要求分类器判断用户输入安全性、助手响应安全性以及有害内容类别,部分示例还涵盖主题跟随分类。训练集输出包含生成推理痕迹和最终标签,而测试集和验证集仅包含最终标签。数据集采用混合数据收集方法,结合了真实或先前收集/整理的数据与合成生成/增强的数据。具体数据来源包括Nemotron安全防护数据集v3、Nemotron内容安全推理数据集、Nemotron VLM数据集v2、Wikimedia Commons以及使用Flux.1 Schnell、Qwen3系列等模型生成的合成图像和提示。人工标注由英伟达内部完成,采用单标注员标注加质量保证专家验证的流程。数据集包含约98,000条记录,其中训练集约88,000条,测试集和验证集各约4,800条。数据模态包括约70,000条纯文本记录和约28,000条图像引用记录。数据集总存储大小为136MB,下载包为67MB。数据集格式为列式结构,包含15个字段:prompt(用户输入)、response(助手响应)、input_label(用户输入安全标签)、response_label(响应安全标签)、image_path(图像路径,可为空字符串、相对路径或Wikimedia Commons URL)、violated_categories(违反的安全类别)、language(语言)、dataset_source(数据来源)、provenance(数据溯源)、row_id(行ID)、reasoning_trace(推理痕迹)、task_type(任务类型)以及各字段的来源信息。图像数据通过单独的tarball(images.tar.gz)提供,Wikimedia Commons图像需通过辅助脚本下载。该数据集旨在用于训练、验证和评估内容安全分类器或视觉语言模型,支持用户安全分类、响应安全分类、有害类别标记、拒绝行为评估、多模态内容安全评估和主题跟随分类等任务,但不适用于原始用户界面展示。
The Nemotron 3.5 Content Safety Dataset is a hybrid real/synthetic supervised instruction dataset designed for human-computer interaction content safety classification. Created by NVIDIA in June 2026, the dataset is licensed under CC-BY-4.0 and permits commercial usage. It encompasses single-turn dialogue examples in both plain-text and image-based modalities, wherein each instance requires the classifier to assess the safety of user inputs, the safety of assistant responses, and identify harmful content categories; some instances additionally cover topic following classification tasks. The training set includes both generated reasoning traces and final labels, while the test and validation sets only contain final labels. The dataset employs a hybrid data collection methodology, combining real or previously curated/aggregated data with synthetically generated or augmented data. Specific data sources include the Nemotron Safety Guard Dataset v3, Nemotron Content Safety Reasoning Dataset, Nemotron VLM Dataset v2, Wikimedia Commons, as well as synthetic images and prompts generated by models including Flux.1 Schnell and the Qwen3 series. Manual annotation was conducted internally by NVIDIA, following a workflow of single annotator labeling supplemented by quality assurance expert validation. The dataset comprises approximately 98,000 total records, with around 88,000 in the training split, and roughly 4,800 records each in the test and validation splits. In terms of data modalities, there are approximately 70,000 plain-text records and 28,000 image-referencing records. The total storage size of the dataset is 136 MB, with the downloadable package being 67 MB in size. The dataset adopts a columnar structure and includes 15 fields: prompt (user input), response (assistant response), input_label (safety label for user input), response_label (safety label for assistant response), image_path (image path, which can be an empty string, relative path, or Wikimedia Commons URL), violated_categories (violated safety categories), language, dataset_source (data source), provenance (data provenance), row_id (row ID), reasoning_trace (reasoning trace), task_type (task type), and source information for each individual field. Image data is distributed via a separate tarball (images.tar.gz), and Wikimedia Commons-hosted images require downloading through an auxiliary script. This dataset is intended for training, validating, and evaluating content safety classifiers or vision-language models, supporting tasks including user safety classification, response safety classification, harmful category labeling, refusal behavior assessment, multimodal content safety evaluation, and topic following classification; however, it is not suitable for raw user interface display.
数据集名称
Nemotron 3.5 Content Safety Dataset
数据集描述
这是一个混合了真实数据与合成数据的监督式指令微调数据集,用于对人与助手的交互进行内容安全分类。数据集包含纯文本和基于图像的单轮对话。每个示例要求分类器判断用户安全性、响应安全性和有害类别;子集还覆盖了主题遵循分类。部分训练输出包含生成推理轨迹,而测试和验证输出仅包含最终标签。
数据集所有者
NVIDIA Corporation
创建日期
2026年6月2日
版本
V2
许可协议
Creative Commons Attribution 4.0 International License (CC-BY-4.0)
预期用途
用于训练、验证和评估内容安全分类器或视觉语言模型(VLM),覆盖用户安全分类、响应安全分类、有害类别标记、拒绝行为评估、多模态内容安全评估和主题遵循分类。不适用于面向用户的原始展示。
数据特征
数据集包含以下特征列:
- prompt (string)
- response (string)
- input_label (string)
- response_label (string)
- image_path (string)
- violated_categories (string)
- language (string)
- dataset_source (string)
- provenance (string)
- row_id (string)
- reasoning_trace (string)
- task_type (string)
- prompt_source (string)
- response_source (string)
- image_source (string)
- input_label_source (string)
- response_label_source (string)
数据划分
- 训练集:88,688 条
- 测试集:4,814 条
- 验证集:4,814 条
- 总记录数:约 98,000 条
数据构成
- 纯文本记录:约 70,000 条
- 图像引用记录:约 28,000 条
- 存储大小:67.1 MB(训练集 63 MB,测试集 2 MB,验证集 2 MB)
数据收集方法
混合自动化与合成方法。数据来源包括:
- Nemotron Safety Guard Dataset v3
- Nemotron Content Safety Reasoning Dataset
- Nemotron VLM Dataset v2
- Wikimedia Commons
- 合成生成的图像和提示词
数据标注方法
人工标注。标注流程为:
- 标注员确定安全主题(如脏话、犯罪策划)。
- 从批准来源获取相关图像。
- 为图像开发合适的文本提示。
- 通过后端系统使用三个LLM生成响应。
- 标注员为提示词、图像和模型输出分配安全标签。
数据格式
列式格式,所有安全标签为文本。合成图像以 tarball (images.tar.gz) 提供,Wikimedia Commons 图像以链接形式提供。图像路径 (image_path) 可能为空字符串、相对路径或 Wikimedia Commons URL。
道德考量
NVIDIA 强调可信赖 AI 是共同责任,开发者应确保数据集满足行业和用例要求,并解决潜在的误用问题。质量问题、风险、安全漏洞或 AI 相关顾虑可通过安全漏洞提交页面报告。




