Nemotron-SFT-Safety-v1 is a safety alignment dataset released by NVIDIA in 2025. Combining open-source collected prompts with synthetically generated responses, this dataset can be used for Supervised Fine-Tuning (SFT) to help models reduce harmful outputs, mitigate over-refusal, lower bias, and prevent sensitive content leakage. It aims to enhance the capabilities of large language models in areas such as content safety, jailbreak defense, and safety behavior alignment.

The dataset contains 45,145 samples in JSONL format, primarily consisting of prompts and synthetically generated safe responses. It is licensed under the Creative Commons Attribution 4.0 International license (CC BY 4.0), with supplementary licenses including Apache 2.0 and MIT.

Dataset Composition

• Content Safety (25,536 entries): Sources include the Nemotron Content Safety Dataset v2, the Gretel Safety Alignment Dataset v1, and synthetically generated prompts for sensitive categories, covering risks such as self-harm, violence, and criminal planning.
• Jailbreak Robustness (9,653 entries): Sources include HarmfulTasks, RedTeam 2k, and synthetically generated jailbreak and potential prompt injection data, used to improve the model's ability to resist attacks.
• Over-Refusal Mitigation (7,023 entries): Helps models reduce unnecessary refusals of normal requests by constructing contrasting prompts that are syntactically similar but semantically different.
• Bias Reduction (2,000 entries): Targets biases related to demographic characteristics such as age, reducing the model's tendency to produce discriminatory responses.
• Sensitive Content Leakage Protection (933 entries): Used to lower the risk of the model leaking potentially copyrighted content.