Nemotron Safety Guard Dataset v3 Multilingual

## Overview The Nemotron-Safety-Guard-Dataset-v3 (formerly Nemotron-Content-Safety-Dataset-Multilingual-v1) is a large, high-quality safety dataset designed for training multilingual LLM safety guard models. It contains approximately 514,617 samples spanning 12 languages: English, Arabic, German, Spanish, French, Hindi, Italian, Japanese, Korean, Dutch, Thai, and Mandarin Chinese. The data is JSON-formatted and primarily synthetically generated via NVIDIA's CultureGuard pipeline for content moderation, toxicity detection, and safety classification training. ## Schema - prompt — string — user prompt to be evaluated for safety - response — string — model or assistant response paired with the prompt - prompt_label — string/categorical — safety label assigned to the prompt (safe/unsafe + category) - response_label — string/categorical — safety label assigned to the response - language — string — ISO language code (one of 12 supported languages) - violated_categories — list[string] — safety taxonomy categories violated (if any) - (+ additional metadata columns; see HF dataset card) ## Sources - nvidia/Nemotron-Safety-Guard-Dataset-v3 on Hugging Face — https://huggingface.co/datasets/nvidia/Nemotron-Safety-Guard-Dataset-v3 — License: CC-BY-4.0 - Associated paper: CultureGuard (arXiv:2508.01710) ## Methodology The dataset was primarily synthetically generated using NVIDIA's CultureGuard pipeline, which produces culturally and linguistically adapted safety training data across 12 languages. Prompts and responses are labeled against a safety taxonomy covering toxicity, harmful content, and other unsafe categories. See the linked arXiv paper and dataset card for full pipeline details, including translation, cultural adaptation, and quality filtering steps. ## Known gaps & limitations As a primarily synthetic dataset, distributions may not reflect real-world user prompts; coverage across the 12 languages is not guaranteed to be uniform in either volume or category coverage. Safety taxonomies are subjective and culture-specific — labels reflect NVIDIA/CultureGuard's taxonomy choices. Buyers training guard models should validate empirically on their target deployment distribution, and audit for any leakage against public safety benchmarks (e.g., ToxicChat, BeaverTails, XSTest). ## Intended use & out-of-scope - IS for: training and fine-tuning multilingual LLM safety/guard classifiers, content moderation evaluation, toxicity detection research, multilingual safety alignment. - NOT for: use as a stand-alone safety benchmark without holdout validation; not deduplicated against common safety eval suites — leakage risk if used naively for benchmark training; synthetic generation means it should not be treated as ground truth for real-world user behavior. _Federated dataset: 3 parquet shards, 242.0 MB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: Nemotron Safety Guard Dataset v3 (Multilingual LLM Safety, 12 Languages) NVIDIA's 514K-sample multilingual safety dataset for training LLM safety guard models across 12 languages, generated via the CultureGuard pipeline. CC-BY-4.0.