Fuleee's picture
Update README.md
f950961 verified
metadata
license: mit
task_categories:
  - text-generation
  - fill-mask
language:
  - zh
tags:
  - java
  - spring-boot
  - code-completion
  - peft
  - lora
  - qlora
  - code-dataset
  - instruction-tuning
pretty_name: SpringBoot-Code-Completion-Dataset-zh
size_categories:
  - 100K<n<1M

Spring Boot Code Completion Dataset

概述

这是一个专为 Spring Boot 框架代码补全任务设计的中文优先数据集。通过从 GitHub 上 90 个高质量开源 Spring Boot 项目中提取方法级、类级代码片段及配置文件,构建而成。数据集重点突出 Spring Boot 典型特征(注解驱动、依赖注入、配置绑定等),适用于参数高效微调(PEFT,如 LoRA/QLoRA)下的代码生成研究。

数据集统计

  • 总样本数:81,085 条
  • 训练集:64,868 条
  • 验证集:8,108 条
  • 测试集:8,109 条
  • 优先级 2 样本数(含典型 Spring Boot 特征,如 @RestController、@Service、@Entity 等):19,740 条
  • 优先级 2 占比:24.34%

数据格式(Alpaca 风格)

每条数据为 JSON 对象,包含以下字段:

{
  "instruction": "请在Spring Boot框架中完成以下Java代码""请完成以下Java代码",
  "input": "代码前缀(prefix)",
  "output": "代码补全部分(completion)",
  "source_file": "原始文件路径",
  "priority": 12   // 2 表示含有 Spring Boot 典型特征
}
提取与清洗流程

从 GitHub 采集 90 个典型 Spring Boot 项目
使用 javalang 进行 AST 解析,提取方法/类级代码片段及 application 配置
多维度 Spring Boot 特征检测(注解、接口实现、包路径、类名后缀等)标记 priority
清洗:精确 MD5 去重 + SentenceTransformer + FAISS 相似度去重(阈值 0.95)+ priority 2 优先保护
长度过滤:input > 50 字符,output > 20 字符
划分:80% 训练 / 10% 验证 / 10% 测试

使用方式
Pythonfrom datasets import load_dataset

dataset = load_dataset("您的用户名/springboot-code-completion-dataset")
# 或指定子集
train = dataset["train"]
许可
本数据集仅限学术研究与非商业用途使用。原始代码来源于 GitHub 开源项目,请遵守各项目原许可协议。
引用
如果在论文中使用,请引用本仓库:
text@misc{唯一2026-springboot-dataset,
  author = {唯一},
  title = {Spring Boot Code Completion Dataset},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/您的用户名/springboot-code-completion-dataset}}
}
欢迎反馈与合作!