Nemotron-Safety-30K

# **Nemotron-Safety-30K**

Nemotron-Safety-30K is a modular post-training dataset specifically designed for safety-based model training. This dataset has been retrieved and curated from the larger Llama-Nemotron-Post-Training-Dataset to focus on safety-related training scenarios.

## Dataset Details

- **Size**: 31,426 rows
- **Format**: Parquet
- **License**: CC-BY-4.0
- **File Size**: 14 MB
- **Modalities**: Text
- **Libraries**: Datasets, pandas, Croissant

## Dataset Structure

The dataset contains the following columns:

- **input**: List of conversation inputs with role-based structure
- **output**: String responses with associated lengths
- **category**: Classification category (primarily "safety")
- **generator**: Model generator information (Mixtral-8x22B-Instruct-v0.1 & more)
- **license**: License information (cc-by-4.0)
- **reasoning**: String indicating reasoning approach ("on" or "off")

## Data Format

Each row follows a conversational format with role-based inputs:
```
{
"role": "user",
"content": "[User query or prompt]"
}
```

The outputs provide safety-oriented responses designed for training models to handle potentially sensitive or harmful content appropriately.

## Source

This dataset is derived from the comprehensive Llama-Nemotron-Post-Training-Dataset available at:
https://huggingface.co/datasets/nvidia/Llama-Nemotron-Post-Training-Dataset

## Usage

This dataset is intended for researchers and developers working on:
- Safety alignment in large language models
- Post-training refinement for responsible AI
- Safety evaluation and benchmarking
- Modular training approaches for specific safety scenarios

## License

This dataset is under following licensing:
- CC-BY-4.0 for the dataset content

## Citation

```bibtex
@misc{llama-nemotron-2025,
title={Llama-Nemotron: Efficient Reasoning Models},
author={Nvidia},
year={2025},
howpublished={Hugging Face Datasets},
url={https://huggingface.co/datasets/nvidia/Llama-Nemotron-Post-Training-Dataset}
}
```

# **Nemotron-Safety-30K**

Nemotron-Safety-30K 是一款专为安全导向型模型训练打造的模块化后训练数据集。该数据集从规模更大的 Llama-Nemotron-Post-Training-Dataset 中筛选并整理而来，专门聚焦安全相关的训练场景。

## 数据集详情

- **数据量**：31426 条记录
- **格式**：Parquet
- **许可证**：CC-BY-4.0
- **文件大小**：14 MB
- **模态**：文本
- **依赖库**：Datasets、pandas、Croissant

## 数据集结构

该数据集包含以下列：

- **input**：包含角色化结构的对话输入列表
- **output**：附带长度信息的字符串回复
- **category**：分类类别（主要为“安全”类）
- **generator**：模型生成器信息（如 Mixtral-8x22B-Instruct-v0.1 等）
- **license**：许可证信息（cc-by-4.0）
- **reasoning**：指示推理模式的字符串（取值为“on”或“off”）

## 数据格式

每条数据均采用角色化对话格式，输入示例如下：

{
"role": "user",
"content": "[用户查询或提示词]"
}

该数据集的回复均为安全导向型内容，用于训练模型以妥善处理潜在敏感或有害信息。

## 数据来源

本数据集源自规模完整的 Llama-Nemotron-Post-Training-Dataset，可通过以下链接获取：
https://huggingface.co/datasets/nvidia/Llama-Nemotron-Post-Training-Dataset

## 使用场景

本数据集面向开展以下研究与开发工作的科研人员与开发者：
- 大语言模型（Large Language Model，LLM）的安全对齐
- 负责任人工智能的后训练优化
- 安全评估与基准测试
- 针对特定安全场景的模块化训练方法

## 许可证

本数据集采用以下许可证：
- 数据集内容遵循 CC-BY-4.0 协议

## 引用格式

bibtex
@misc{llama-nemotron-2025,
title={Llama-Nemotron: Efficient Reasoning Models},
author={Nvidia},
year={2025},
howpublished={Hugging Face Datasets},
url={https://huggingface.co/datasets/nvidia/Llama-Nemotron-Post-Training-Dataset}
}