apcl/funcom-java-long

# funcom-java-long

funcom-java-long is a dataset for code summarization presented by Bansal et al. It includes 170k Java methods and summaries for training and 8,192 Java methods in test set. This dataset is extracted from jm52m and is cleaned and deduplicated. The example prompt for fine-tuning is as follows.

- TDAT: \<method code\> COMMENT: \<comment\> \<!endofdoc\>

### funcom_test.tar.gz -- a compressed directory for test set with the text inside this folder whose filename = function id
### funcom_train.tar.gz -- a compressed directory for training set with the text inside this folder whose filename = function id
### funcom_val.tar.gz -- a compressed directory for validation set with the text inside this folder whose filename = function id
### train.bin; val.bin -- bin files for training in the format above.
### test.bin; bin file with the test data in the format above.

原始信息汇总

funcom-java-long 数据集概述

数据集简介

• 名称：funcom-java-long
• 目的：用于代码摘要生成
• 贡献者：Bansal et al.
• 数据来源：从jm52m数据集提取并经过清洗和去重处理

数据组成

• 训练集：包含170,000个Java方法及其摘要
• 测试集：包含8,192个Java方法

数据集文件

• 测试集：funcom_test.tar.gz，包含以函数ID命名的文本文件
• 训练集：funcom_train.tar.gz，包含以函数ID命名的文本文件
• 验证集：funcom_val.tar.gz，包含以函数ID命名的文本文件
• 训练数据二进制文件：train.bin 和 val.bin
• 测试数据二进制文件：test.bin

数据格式

• 示例提示：TDAT: <method code> COMMENT: <comment> <!endofdoc>