MOSAIC-agentic-3m
收藏资源简介:
该数据集由代尔夫特理工大学等机构联合构建,收录了约11万条开源GitHub pull requests及其关联元数据,涵盖提交、评论、审查记录等问题追踪信息,涉及数百万行源代码。数据通过挖掘五个主流编码代理(包括OpenAI Codex、GitHub Copilot等)的协作痕迹构建,重点捕捉AI代理与人类开发者的交互模式。数据集旨在研究自主编码代理对代码质量、团队协作及软件可维护性的影响,为AI原生软件开发提供实证分析基础。
This dataset was jointly constructed by Delft University of Technology and other institutions. It contains approximately 110,000 open-source GitHub Pull Requests and their associated metadata, covering issue tracking information such as commits, comments and review records, and involving millions of lines of source code. The dataset was built by mining collaboration traces from five mainstream coding agents, including OpenAI Codex, GitHub Copilot and others, with a focus on capturing the interaction patterns between AI Agents and human developers. This dataset aims to study the impact of autonomous coding agents on code quality, team collaboration and software maintainability, providing an empirical analysis foundation for AI-native software development.
数据集概述
基本信息
- 数据集名称: MOSAIC-agentic-3m
- 发布者: AISE-TUDelft
- 许可证: GPL-3.0
- 地址: https://huggingface.co/datasets/AISE-TUDelft/MOSAIC-agentic-3m
数据集结构
数据集包含多个配置,每个配置对应一种贡献者类型和一种活动类型。
配置概览
配置名称遵循 {活动类型}_{贡献者类型} 的格式。
活动类型
- Comments (评论)
- Commits (提交)
- Issues (议题)
- PullRequests (拉取请求)
贡献者类型
- Claude
- Codex
- Copilot
- Devin
- Human
- Jules
各配置详情
Comments 配置组
包含 Comments_Claude, Comments_Codex, Comments_Copilot, Comments_Devin, Comments_Human, Comments_Jules。
- 通用特征:
id: 字符串pr_id: 字符串url: 字符串body: 字符串created_at: 字符串is_minimized: 布尔值minimized_reason: 字符串last_edited_at: 字符串published_at: 字符串updated_at: 字符串author: 结构体 (包含id,login,name,typename,url)
- 数据规模:
- Comments_Claude: 22,329 个样本,55.86 MB
- Comments_Codex: 3,693 个样本,4.62 MB
- Comments_Copilot: 26,664 个样本,32.99 MB
- Comments_Devin: 27,518 个样本,25.81 MB
- Comments_Human: 18,559 个样本,23.91 MB
- Comments_Jules: 5,700 个样本,6.18 MB
Commits 配置组
包含 Commits_Claude, Commits_Codex, Commits_Copilot, Commits_Devin, Commits_Human, Commits_Jules。
- 通用特征:
id: 字符串sha: 字符串pr_id: 字符串url: 字符串committed_date: 字符串additions: 整数deletions: 整数authored_date: 字符串message_body: 字符串message_headline: 字符串author_count: 整数committer: 结构体 (包含email,name)changed_files: 整数authors: 列表 (元素为包含email,name的结构体)
- 数据规模:
- Commits_Claude: 82,755 个样本,78.33 MB
- Commits_Codex: 27,530 个样本,13.02 MB
- Commits_Copilot: 69,896 个样本,41.97 MB
- Commits_Devin: 51,641 个样本,45.60 MB
- Commits_Human: 102,037 个样本,54.51 MB
- Commits_Jules: 41,032 个样本,39.45 MB
Issues 配置组
包含 Issues_Claude, Issues_Codex, Issues_Copilot, Issues_Devin, Issues_Human, Issues_Jules。
- 通用特征:
id: 字符串pr_id: 字符串url: 字符串title: 字符串body: 字符串created_at: 字符串locked: 布尔值或空值number: 整数state: 字符串tracked_issues_count: 整数或空值label_count: 整数last_edited_at: 字符串published_at: 字符串updated_at: 字符串issue_type: 结构体 (包含description,name)labels: 列表 (元素为包含description,name的结构体)state_reason: 字符串author: 结构体 (包含id,login,name,typename,url)pr_ids: 空值prs_closing_issue: 整数
- 数据规模:
- Issues_Claude: 4,052 个样本,8.37 MB
- Issues_Codex: 45 个样本,0.06 MB
- Issues_Copilot: 9,744 个样本,18.04 MB
- Issues_Devin: 294 个样本,0.50 MB
- Issues_Human: 1,973 个样本,2.59 MB
- Issues_Jules: 2,185 个样本,4.43 MB
PullRequests 配置组
包含 PullRequests_Claude, PullRequests_Codex (特征列表不完整,详情页中可能未完全展示)。
- 部分通用特征 (基于 PullRequests_Claude):
id: 字符串title: 字符串url: 字符串number: 整数body: 字符串state: 字符串created_at: 字符串is_draft: 布尔值changed_files: 整数is_cross_repository: 布尔值locked: 布尔值is_in_merge_queue: 空值additions: 整数deletions: 整数author: 结构体 (包含id,login,name,typename,url)label_count: 整数base_repository: 结构体 (包含id,name,url)head_repository: 结构体 (包含id,name,url)timeline_count: 整数merged_at: 字符串closed_at: 字符串updated_at: 字符串last_edited_at: 字符串published_at: 字符串review_decision: 字符串head_ref_name: 字符串head_ref_oid: 字符串timeline_items: 空值base_ref_name: 字符串base_ref_oid: 字符串comments_count: 整数reviews_count: 整数commits_count: 整数files: 列表 (元素为包含additions,change_type,deletions,path的结构体)assignees_count: 整数closing_issues_count: 整数author_association: 字符串labels: 列表 (元素为包含description,name的结构体)active_lock_reason: 字符串
- 数据规模 (基于 PullRequests_Claude):
- PullRequests_Claude: 19,148 个样本,60.91 MB
数据规模总结
- 数据集包含多种软件开发活动记录,按贡献者类型分类。
- 总样本数量为各配置样本数之和。
- 所有配置均仅包含训练集 (
trainsplit)。

- 1Investigating Autonomous Agent Contributions in the Wild: Activity Patterns and Code Change over Time代尔夫特理工大学; 加州大学戴维斯分校; GitHub · 2026年



