Voidreaper2026/coding-master-dataset

Coding Master Dataset 是一个大规模编码指令调优数据集，采用 ShareGPT 对话格式，通过汇编多个开源数据源并经过去重处理构建而成。该数据集包含 766,987 条记录，以 JSONL 格式存储，遵循 Apache 2.0 许可证。它专为文本生成和问答任务设计，适用于代码相关的指令调优和微调，覆盖 Python、JavaScript、Go 等编程语言，规模在 10 万到 100 万条之间。数据来源包括 CodeX-2M-Thinking、python-code-dataset-500k、StackPulse 高质量子集、CodeFeedback-Filtered-Instruction、secure_programming_dpo 和 Go 编码 JSONL 文件，确保了多样性和覆盖面。数据集结构包含 conversations（ShareGPT 格式的人机对话轮次）、source（来源数据集）、source_id（唯一标识符）和 metadata_json（附加元数据）等字段，便于使用和扩展。A large-scale coding instruction-tuning dataset in ShareGPT conversational format, assembled from multiple open sources and deduplicated. It contains 766,987 records in JSONL format, licensed under Apache 2.0, and is designed for text-generation and question-answering tasks, suitable for code-related instruction tuning and fine-tuning across programming languages like Python, JavaScript, and Go, with a size category of 100K<n<1M. Sources include CodeX-2M-Thinking, python-code-dataset-500k, StackPulse high-quality subset, CodeFeedback-Filtered-Instruction, secure_programming_dpo, and Go coding JSONL files, ensuring diversity and coverage. The schema features fields such as conversations (ShareGPT format human/gpt turns), source (origin dataset), source_id (unique record identifier), and metadata_json (additional metadata) for ease of use and extensibility.