gss1147/GOD_Coder_Complete_DataSet

--- language: - en license: other license_name: within-us-ai-custom-dataset-license pretty_name: GOD_Coder_Complete_DataSet size_categories: - 100K<n<1M task_categories: - text-generation - question-answering - text-classification tags: - code - coding - software-engineering - instruction-tuning - sft - ai-coding - complete-project-coding - repository-patching - debugging - dependency-resolution - full-stack-engineering - code-review - dataset annotations_creators: - machine-generated - expert-generated language_creators: - machine-generated multilinguality: - monolingual source_datasets: - original viewer: false --- # GOD_Coder_Complete_DataSet ## Subtitle A large-scale complete-project coding dataset by **gss1147 / WithIn Us AI**, built to train language models into stronger professional software-engineering assistants. ## Dataset Summary **GOD_Coder_Complete_DataSet** is a large synthetic supervised fine-tuning dataset designed to help turn a general language model into a **professional complete-project AI coder**. The dataset focuses on teaching models how to: - diagnose realistic repository issues - patch broken code with production-ready fixes - write and repair tests - handle dependency and migration failures - reason across full software stacks - solve advanced coding-logic problems - behave more like a senior engineer on complete software projects This dataset was created by **gss1147** under **WithIn Us AI**. ## Creator - **Creator:** gss1147 - **Organization / Brand:** WithIn Us AI - **Dataset Concept, Design, Structure, and Packaging:** WithIn Us AI - **Primary Author:** gss1147 ## License This dataset uses the **WithIn Us AI Custom Dataset License**. ## Dataset Purpose The purpose of this dataset is to provide a strong supervised fine-tuning resource for training coding-capable LLMs toward: - complete software-project reasoning - professional engineering behavior - multi-file patch generation - debugging and issue resolution - test-backed implementation quality - dependency-aware coding - rollout-safe software delivery - increasingly advanced coding logic This dataset is intended for researchers, model builders, and fine-tuning practitioners who want a model that behaves more like a **real software engineer**, not just a code autocompleter. ## Supported Tasks This dataset is suitable for: - supervised fine-tuning - instruction tuning - coding assistant specialization - software-engineering behavior shaping - repository issue repair - debugging assistance - dependency resolution training - software delivery planning - code review improvement - complete-project coding workflows ## Dataset Structure The dataset is organized into **7 major subject groups**, each containing **25,000 examples**, for a total of **175,000 rows**. ### Subject Groups 1. **AI Coding** 2. **AI Dependency Coding** 3. **AI Coding Stacks** 4. **AI Software Development** 5. **AI Coding Logic Master** 6. **AI Coding Logic Legendary** 7. **AI Coding Logic God** ### Total Size - **Total examples:** 175,000 - **Train examples:** 171,500 - **Validation examples:** 3,500 ## Data Format Each example is stored in **chat-format JSONL** and includes: - `id` - `subject` - `subject_title` - `tier` - `language` - `framework` - `stack` - `domain` - `topic` - `task_type` - `split` - `freshness_bucket` - `source_grounding` - `messages` - `artifacts` - `labels` ### Example Schema ```json { "id": "ai_coding-00001-abcdef1234567890", "subject": "ai_coding", "subject_title": "AI Coding", "tier": "hard", "language": "Python", "framework": "FastAPI", "stack": ["FastAPI", "PostgreSQL", "Redis", "Celery", "pytest", "Docker"], "domain": "auth service", "topic": "JWT refresh token rotation", "task_type": "repo_issue_patch", "split": "train", "freshness_bucket": "synthetic_transformed_post_2025_style", "source_grounding": { "kind": "synthetic_transformed_repo_task", "license_ok": true, "provenance_note": "Synthetic training example designed for coding-model SFT and labeled as synthetic." }, "messages": [ { "role": "system", "content": "You are a production-grade software engineer. Return a correct, secure, complete, test-backed solution with concise reasoning and no placeholders." }, { "role": "user", "content": "Repository domain: auth service..." }, { "role": "assistant", "content": "Diagnosis... implementation... tests... verification..." } ], "artifacts": { "verification_commands": ["pytest -q", "ruff check ."], "requires_tests": true, "format": "chat_sft" }, "labels": { "correctness": 1, "security": 1, "production_ready": 1, "test_quality": 1, "complete_project_focus": 1 } } Languages Covered The dataset includes tasks across multiple coding and infrastructure languages, including: • Python • TypeScript • JavaScript • Go • Rust • Java • C# • C++ • SQL • Bash • YAML Content Overview The dataset emphasizes production-style software engineering. It includes examples involving: • bug fixing • feature implementation • code review correction • API design • dependency resolution • version migration repair • lockfile and reproducibility debugging • full-stack issue handling • rollout-safe software delivery • incident remediation • concurrency and logic debugging • performance bottleneck repair • multi-file patching • security hardening • observability-aware engineering Data Generation Method This dataset was created as a synthetic structured coding dataset for fine-tuning and instruction-tuning purposes. The generation process focused on: • professional software-engineering style prompts • complete implementation responses • test-backed solutions • production-oriented reasoning • multi-stack coverage • advanced logic difficulty bands • complete-project engineering behavior Examples were designed to reflect realistic repository and engineering scenarios while remaining clearly labeled as synthetic. Why This Dataset Exists Many coding datasets over-focus on: • short single-function code tasks • toy algorithm problems • incomplete snippets • beginner-level instruction pairs GOD_Coder_Complete_DataSet was created to push beyond that by training models on: • complete-project coding behavior • software-engineering decision quality • professional debugging patterns • multi-layer issue resolution • deployment-safe thinking • engineering-grade patch quality Intended Use This dataset is intended for: • full-model fine-tuning • instruction tuning • coding model specialization • research into software-engineering-capable LLMs • training models that can operate more effectively in repository-style workflows It is especially relevant for users building: • coding copilots • patch-generation systems • engineering support agents • code-review assistants • debugging assistants • full-stack project agents Recommended Training Uses Recommended uses include: • supervised fine-tuning on chat-formatted LLMs • continued instruction tuning for coding behavior • staged curriculum learning across difficulty tiers • subject-wise training by shard • multi-phase training where foundational coding precedes advanced logic tiers Suggested Progression 1. AI Coding 2. AI Dependency Coding 3. AI Coding Stacks 4. AI Software Development 5. AI Coding Logic Master 6. AI Coding Logic Legendary 7. AI Coding Logic God Source Data • Source Type: Original dataset created by WithIn Us AI • Primary Creator: gss1147 • Dataset Design: WithIn Us AI • Origin: Synthetic and structured software-engineering task generation Data Splits • Train: 171,500 • Validation: 3,500 The split is tracked using the split field inside each example. Dataset Strengths • large-scale • complete-project focus • professional engineering framing • multi-language coverage • test-backed outputs • multi-subject structure • strong software-development emphasis • suited for coding-model specialization • useful for curriculum-based fine-tuning Dataset Limitations • synthetic rather than extracted from real private repositories • does not guarantee novelty against all historic model pretraining corpora • should be combined with careful evaluation • should ideally be paired with held-out benchmark testing • should not be treated as a substitute for licensed real-world patch datasets where available Bias, Risks, and Safety Quality Philosophy The dataset was designed around these principles: • no placeholders • complete answers • production-ready orientation • secure-by-default thinking • tests included as a training signal • full-project engineering mindset • patch and verification awareness Citation BibTeX @dataset{gss1147_god_coder_complete_dataset_2026, author = {gss1147 and WithIn Us AI}, title = {GOD_Coder_Complete_DataSet}, year = {2026}, publisher = {Hugging Face}, note = {Synthetic supervised fine-tuning dataset for professional complete-project AI coding} } Acknowledgment GOD_Coder_Complete_DataSet was created by gss1147 under WithIn Us AI as part of a broader effort to build stronger open coding-focused AI systems with professional software-engineering behavior. Here is the only YAML fix that mattered: ```yaml license: other license_name: within-us-ai-custom-dataset-licenselanguage: - 英语 license: 其他 license_name: within-us-ai-custom-dataset-license pretty_name: GOD_Coder_Complete_DataSet size_categories: - 100,000 < 样本量 < 1,000,000 task_categories: - 文本生成 - 问答 - 文本分类 tags: - 代码 - 编码开发 - 软件工程 - 指令微调（Instruction Tuning） - 监督微调（Supervised Fine-Tuning, SFT） - AI编码开发 - 完整项目编码 - 代码仓库补丁修复 - 调试 - 依赖项解析 - 全栈工程 - 代码审查 - 数据集 annotations_creators: - 机器生成 - 专家生成 language_creators: - 机器生成 multilinguality: - 单语言 source_datasets: - 原创数据集 viewer: 禁用 # GOD_Coder_Complete_DataSet ## 副标题 由**gss1147 / WithIn Us AI**打造的大规模完整项目编码数据集，旨在将大语言模型（Large Language Model, LLM）训练为更专业的软件工程辅助工具。 ## 数据集概述 **GOD_Coder_Complete_DataSet**是一款大规模合成式监督微调数据集，旨在将通用大语言模型打造为**专业的完整项目级AI编码助手**。 本数据集聚焦于赋能模型掌握以下能力： - 诊断真实代码仓库中的实际问题 - 为存在缺陷的代码提供可直接用于生产环境的补丁修复方案 - 编写并修复测试用例 - 处理依赖项与版本迁移故障 - 完成全软件栈的逻辑推理 - 解决高阶编码逻辑问题 - 在完整软件项目开发中展现资深软件工程的思维与行为模式 本数据集由**gss1147**在**WithIn Us AI**主导开发。 ## 创作者信息 - **创作者**：gss1147 - **所属机构/品牌**：WithIn Us AI - **数据集概念、设计、结构与封装**：WithIn Us AI - **主要作者**：gss1147 ## 许可证 本数据集采用**WithIn Us AI 自定义数据集许可证**。 ## 数据集用途 本数据集旨在为具备编码能力的大语言模型提供优质的监督微调资源，助力模型实现以下目标： - 完整软件项目逻辑推理 - 专业软件工程行为模式 - 多文件补丁生成 - 缺陷调试与问题解决 - 基于测试的实现质量保障 - 感知依赖项的编码开发 - 支持安全上线的软件交付 - 高阶编码逻辑能力 本数据集面向希望打造更贴近**真实软件工程师**（而非仅代码自动补全工具）的研究人员、模型开发者与微调从业者。 ## 支持任务场景 本数据集适用于以下场景： - 监督微调 - 指令微调 - 编码助手专业化训练 - 软件工程行为模式塑造 - 代码仓库问题修复 - 调试辅助 - 依赖项解析训练 - 软件交付规划 - 代码审查优化 - 完整项目编码工作流适配 ## 数据集结构 本数据集分为**7大主题组**，每组包含25,000条样本，总计175,000条数据记录。 ### 主题组 1. **AI编码开发** 2. **AI依赖项编码开发** 3. **AI全栈编码开发** 4. **AI软件工程开发** 5. **AI编码逻辑精通** 6. **AI编码逻辑大师** 7. **AI编码逻辑宗师** ### 总规模 - **总样本数**：175,000 - **训练集样本数**：171,500 - **验证集样本数**：3,500 ## 数据格式 每条样本以**对话格式JSONL**存储，包含以下字段： - `id` - `subject` - `subject_title` - `tier` - `language` - `framework` - `stack` - `domain` - `topic` - `task_type` - `split` - `freshness_bucket` - `source_grounding` - `messages` - `artifacts` - `labels` ### 示例 schema json { "id": "ai_coding-00001-abcdef1234567890", "subject": "ai_coding", "subject_title": "AI Coding", "tier": "hard", "language": "Python", "framework": "FastAPI", "stack": ["FastAPI", "PostgreSQL", "Redis", "Celery", "pytest", "Docker"], "domain": "auth service", "topic": "JWT refresh token rotation", "task_type": "repo_issue_patch", "split": "train", "freshness_bucket": "synthetic_transformed_post_2025_style", "source_grounding": { "kind": "synthetic_transformed_repo_task", "license_ok": true, "provenance_note": "为编码模型监督微调设计的合成训练样本，标注为合成数据。" }, "messages": [ { "role": "system", "content": "你是一名生产级软件工程师。请返回正确、安全、完整且基于测试的解决方案，附带简洁推理过程，无任何占位符。" }, { "role": "user", "content": "代码仓库业务领域：认证服务..." }, { "role": "assistant", "content": "缺陷诊断...代码实现...测试用例...验证方案..." } ], "artifacts": { "verification_commands": ["pytest -q", "ruff check ."], "requires_tests": true, "format": "chat_sft" }, "labels": { "correctness": 1, "security": 1, "production_ready": 1, "test_quality": 1, "complete_project_focus": 1 } } ## 覆盖语言 本数据集覆盖多种编码与基础设施语言，包括： • Python • TypeScript • JavaScript • Go • Rust • Java • C# • C++ • SQL • Bash • YAML ## 内容概览 本数据集以生产级软件工程为核心导向，涵盖以下场景： • 缺陷修复 • 功能开发 • 代码审查修正 • API设计 • 依赖项解析 • 版本迁移修复 • 锁文件与可复现性调试 • 全栈问题处理 • 安全上线的软件交付 • 事件处置 • 并发与逻辑调试 • 性能瓶颈修复 • 多文件补丁修复 • 安全加固 • 可观测性导向的工程开发 ## 数据生成方法 本数据集为面向微调与指令微调场景打造的合成式结构化编码数据集。 数据集生成过程聚焦于以下要点： • 专业软件工程风格的提示词 • 完整的实现类响应 • 基于测试的解决方案 • 生产导向的逻辑推理 • 多技术栈覆盖 • 分阶的高阶逻辑难度 • 完整项目级工程行为模式 所有样本均设计为还原真实代码仓库与工程场景，同时明确标注为合成数据集。 ## 本数据集的开发初衷 现有多数编码数据集过度聚焦于： • 简短的单函数代码任务 • 玩具级算法问题 • 不完整的代码片段 • 入门级指令对 **GOD_Coder_Complete_DataSet**旨在突破上述局限，通过以下场景训练模型： • 完整项目级编码行为 • 软件工程决策能力 • 专业调试模式 • 多层级问题解决 • 部署安全思维 • 工程级补丁质量 ## 预期应用场景 本数据集面向以下应用场景： • 全模型微调 • 指令微调 • 编码模型专业化训练 • 面向具备软件工程能力的大语言模型研究 • 训练可在代码仓库式工作流中高效运行的模型 尤其适用于开发以下工具的用户： • 编码助手 • 补丁生成系统 • 工程支持智能体 • 代码审查助手 • 调试助手 • 全栈项目智能体 ## 推荐训练使用方式 推荐的训练使用方式包括： • 面向对话格式大语言模型的监督微调 • 针对编码行为的持续指令微调 • 基于难度分层的分阶段课程学习 • 按主题分片的专项训练 • 先基础编码、后高阶逻辑的多阶段训练 ## 推荐进阶路径 1. AI编码开发 2. AI依赖项编码开发 3. AI全栈编码开发 4. AI软件工程开发 5. AI编码逻辑精通 6. AI编码逻辑大师 7. AI编码逻辑宗师 ## 源数据信息 • 数据源类型：由WithIn Us AI原创打造的数据集 • 主要创作者：gss1147 • 数据集设计：WithIn Us AI • 数据来源：合成式结构化软件工程任务生成 ## 数据集划分 • 训练集：171,500条 • 验证集：3,500条 数据集划分通过每条样本内的`split`字段进行标识。 ## 数据集优势 • 大规模规模 • 聚焦完整项目开发 • 专业工程化框架 • 多语言覆盖 • 基于测试的输出结果 • 多主题结构 • 强化软件开发场景 • 适配编码模型专业化训练 • 适用于课程式微调训练 ## 数据集局限性 • 数据为合成生成，而非从真实私有仓库提取 • 无法保证相较于所有历史模型预训练语料的新颖性 • 需结合严谨的评估方案使用 • 建议配合预留基准测试集使用 • 若有可用的授权真实世界补丁数据集，不应将本数据集作为其替代品 ## 偏见、风险与安全 ## 质量理念 本数据集围绕以下原则设计： • 无占位符内容 • 完整答案 • 生产就绪导向 • 默认安全思维 • 将测试用例作为训练信号 • 全项目工程思维 • 补丁与验证意识 ## 引用信息 ### BibTeX格式引用 bibtex @dataset{gss1147_god_coder_complete_dataset_2026, author = {gss1147 and WithIn Us AI}, title = {GOD_Coder_Complete_DataSet}, year = {2026}, publisher = {Hugging Face}, note = {面向专业完整项目级AI编码的合成式监督微调数据集} } ## 致谢 **GOD_Coder_Complete_DataSet**由gss1147在WithIn Us AI主导开发，旨在推动打造具备专业软件工程行为能力的开源编码导向AI系统。 以下为唯一有效的YAML许可证配置： yaml license: other license_name: within-us-ai-custom-dataset-license