SoftAge-AI/rlhf-ranking_dataset

---
license: mit
language:
- en
---

# RLHF Response Ranking Dataset

## Description

This dataset supports research in Response Ranking for Large Language Models (RLHF) in the CODE & STEM domain.

It contains 500 prompt-response pairs, each with the following data attributes:

- M_Id & S.No.: Unique identifier for the prompt-response pair.
- Prompt: The original query or problem statement.
- Response 1 & 2: Responses generated by different language models.
- prompt_type: Category of the prompt (e.g., mathematical equation, coding problem).
- Preference: Indicates which response is considered better (1 or 2).
- Remark: Additional information about the ranking decision.
- Safety labels (all Y/N):
- Fails to follow instructions
- Contains sexual content
- Contains violent content
- Encourages harmful behavior
- Expresses moral judgment
- Gives harmful advice

## Dataset Source

This dataset is curated by the delivery team @SoftAge

## Limitations and Biases

- This dataset might not capture the full diversity of CODE & STEM problems and response qualities.
- Preference labels and safety ratings might reflect the inherent biases of human annotators or domain experts.

## Potential Uses

• Training and analysing RLHF models for generating informative and safe responses in the CODE & STEM domain.
• Identifying areas for improvement in language models.
• Developing new metrics and methods for RLHF in different domains.

原始信息汇总

RLHF Response Ranking Dataset

描述

该数据集支持在CODE & STEM领域中对大型语言模型（RLHF）的响应排序研究。

它包含500个提示-响应对，每个对具有以下数据属性：

• M_Id & S.No.: 提示-响应对的唯一标识符。
• Prompt: 原始查询或问题陈述。
• Response 1 & 2: 由不同语言模型生成的响应。
• prompt_type: 提示的类别（例如，数学方程，编程问题）。
• Preference: 指示哪个响应被认为是更好的（1或2）。
• Remark: 关于排序决策的附加信息。
• Safety labels (all Y/N):
  • 未能遵循指令
  • 包含性内容
  • 包含暴力内容
  • 鼓励有害行为
  • 表达道德判断
  • 给出有害建议

数据集来源

该数据集由SoftAge的交付团队策划。

限制和偏见

• 该数据集可能无法捕捉CODE & STEM问题和响应质量的全部多样性。
• 偏好标签和安全评级可能反映了人类标注者或领域专家的固有偏见。

潜在用途

• 训练和分析RLHF模型，以在CODE & STEM领域生成信息丰富且安全的响应。
• 识别语言模型改进的领域。
• 开发不同领域中RLHF的新指标和方法。