minfeng-ai/leetcode_preference

该数据集用于支持基于Direct Preference Optimization（DPO）的实验，该优化方法在论文《Direct Preference Optimization: Your Language Model is Secretly a Reward Model》中有详细介绍。数据集包含由CodeLLaMA-7b生成的每道LeetCode问题的两个独特解决方案，这些解决方案由人类专家根据其准确性、效率和可读性进行评估和排名。数据集中的每一行包含以下字段：id（唯一标识符）、title（问题官方标题）、description（详细提示）、difficulty（难度分为简单、中等和困难三个等级）、version1（与问题相关的初始AI生成的代码片段）、version2（与同一问题相关的第二AI生成的代码片段）和preference（人类评估的标签，指示Version1和Version2之间的首选代码片段）。

原始信息汇总

Dataset Card for LeetCode Preference

Dataset Description

Dataset Summary

• Purpose: Facilitates experiments utilizing Direct Preference Optimization (DPO).
• Source Paper: Direct Preference Optimization: Your Language Model is Secretly a Reward Model
• Code Generation: Code pairings crafted by CodeLLaMA-7b for every LeetCode question.
• Evaluation: Human experts evaluate and rank two unique solutions per question based on accuracy, efficiency, and readability.

Usage

python from datasets import load_dataset

dataset = load_dataset("minfeng-ai/leetcode_preference")

Data Fields

• id: Distinct identifier for each LeetCode question.
• title: Official title of the LeetCode question.
• description: Detailed prompt for the question.
• difficulty: Question complexity categorized as Easy, Medium, or Hard.
• version1: Initial AI-generated code snippet.
• version2: Secondary AI-generated code snippet.
• preference: Human-assessed label indicating the preferred code snippet between Version1 and Version2.