LukasSonn/DoxygenStrings-Long

---
license: apache-2.0
---
# Dataset Info C++ + Natural Description -> Doxygen Documentation

This dataset was created for my bachelors thesis investigating how LLMs can be fine-tuned to generate doxygen documentation. It was created by using the “Source code analysis dataset”
by Gelman, Banjo Obayomi, Jessica Moore und David Slater (doi: 10.1016/j.dib.2019.104712).

The following SQL-Statement was used to grab raw data from the dataset:
```
SELECT * FROM all_data
WHERE LENGTH(comment) > 300 and LENGTH(code) > 100 AND LENGTH(code) < 80
AND code NOT LIKE '%//%' AND code NOT LIKE '%/*%' AND code NOT LIKE '%*/%'
AND filename LIKE '%.cpp%'
LIMIT 12000
```

After selecting the Data Code LLaMa Instruct 34B is tasked to combine the human-written description of the functionality with the function code into a Doxygen-Comment. Any results which included the sample doxygen string or no doxygen string at all where filtered from the set.

原始信息汇总

数据集信息：C++代码与自然描述生成Doxygen文档

数据集来源

• 数据集基于Gelman, Banjo Obayomi, Jessica Moore和David Slater的“源代码分析数据集”（doi: 10.1016/j.dib.2019.104712）创建。

数据选择标准

• 使用SQL语句从源数据集中筛选数据： sql SELECT * FROM all_data WHERE LENGTH(comment) > 300 and LENGTH(code) > 100 AND LENGTH(code) < 80 AND code NOT LIKE %//% AND code NOT LIKE %/% AND code NOT LIKE %/% AND filename LIKE %.cpp% LIMIT 12000

数据处理

• 使用Code LLaMa Instruct 34B模型将人工编写的功能描述与函数代码结合，生成Doxygen注释。
• 过滤掉包含示例Doxygen字符串或无Doxygen字符串的结果。