Zovda/dm0012-vibe-coding-samples

DM0012 · Vibe Coding Repository-Level Task Samples / Vibe Coding 仓库级任务样例

Item / 项目	Value / 值
Dataset ID / 数据集编号	DM0012
Files / 数据文件	1
Size / 数据体积	29.36 KB
Provided by / 提供方	Zovda
Contact / 联系邮箱	lucas@zovda.com

Overview / 简介

Repository-level task corpus for coding-agent (SWE-Agent) training and evaluation: each package targets a real GitHub issue and bundles a runnable Docker sandbox, task description, reference patches and automatic grading scripts.

面向代码智能体（SWE-Agent）训练与评测的仓库级任务语料：每个任务包对应一个真实开源仓库的 GitHub Issue，含可运行 Docker 沙箱、任务描述、参考补丁与自动验收脚本。

What's inside / 数据内容

• manifest.json: the public task-package manifest listing name, size, date, org, repo and issue id for each package.
• manifest.json：随样例库公开的任务包清单，逐条记录包名、大小、日期、org、repo 与 issue 编号。
• Full packages (<org>__<repo>-<issueId>.zip) are held in a private delivery store; about 42 entries per package — docker/<issueId>/ (OCI-layout sandbox image) plus json/<org__repo-issueId>/ (task description, code.patch, model.patch, final.diff, test.patch, check_delivery.py, grade.sh, dockerfile, setup_env.sh, setup_repo.sh, pin_by_date.py).
• 完整任务包（<org>__<repo>-<issueId>.zip）存放于私有交付目录，每包约 42 个条目，结构为：docker/<issueId>/（OCI 布局沙箱镜像）与 json/<org__repo-issueId>/（任务描述、code.patch、model.patch、final.diff、test.patch、check_delivery.py、grade.sh、dockerfile、setup_env.sh、setup_repo.sh、pin_by_date.py）。
• Contact lucas@zovda.com to obtain packages.
• 取数请联系 lucas@zovda.com。

Specifications / 规格

• Packaging: Docker (OCI layout) + JSON, two layers
• 打包：Docker（OCI 布局）+ JSON 双层
• Naming: <org>__<repo>-<issueId>, traceable to the real issue
• 命名：<org>__<repo>-<issueId>，issueId 可回溯真实 Issue
• Acceptance: grade.sh auto-grades inside the sandbox; check_delivery.py verifies delivery completeness
• 验收：grade.sh 在沙箱内自动评分，check_delivery.py 校验交付完整性

Suitable tasks / 适用任务

• Code-model post-training (SFT): paired "issue description → reference patch" samples
• 代码模型后训练（SFT）：「Issue 描述 → 参考补丁」成对样本
• Reinforcement learning / agent training: sandbox + auto-grading provide reward signals
• 强化学习 / Agent 训练：沙箱 + 自动评分提供奖励信号
• End-to-end coding-agent evaluation: issue-resolution pass rate
• 代码智能体端到端评测：Issue 修复通过率

Files / 文件清单

#	File	Size
1	vibeCoding/manifest.json	29.36 KB

Usage / 使用方式

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Zovda/dm0012-vibe-coding-samples",
    repo_type="dataset",
    local_dir="./dm0012-vibe-coding-samples",
)

Rights and disclaimer / 授权与声明

Tasks are built from real issues in public open-source repositories, which retain their original open-source licences. This manifest demonstrates task organisation; packages are licensed on request.

任务基于公开开源仓库的真实 Issue 构建，仓库代码遵循其原始开源许可。本清单用于展示任务组织方式，任务包按需授权交付。

This dataset is shared to demonstrate data organisation, technical specifications and delivery structure. For formal licensing, large-scale access or custom production, contact lucas@zovda.com.

本数据集仅用于展示数据组织方式、技术规格与交付结构。正式授权、规模化取数与定制生产请联系 lucas@zovda.com。

About Zovda / 关于 Zovda

Zovda builds, cleans, annotates and packages training data across video, audio, images, e-commerce image-text pairs, exam question banks, embodied AI and coding-agent corpora.

Zovda 提供多模态与文本训练语料的数据采集、清洗、标注与结构化交付服务，覆盖视频、音频、图像、电商图文、K12 与职业题库、具身智能、代码智能体等方向。

Contact / 联系方式: lucas@zovda.com