APIBench is the benchmark for evaluating the performance of API recommendation approaches released in paper "Revisiting, Benchmarking and Exploring API Recommendation: How Far Are We?".

As GitHub does not hold large datasets, you can download the benchmark dataset at Zenodo.

Currently APIBench contains two sub-dataset for evaluating the performance of query-based and code-based API recommendation approaches, namely APIBench-Q and APIBench-C. Each sub-dataset has a Java version and a Python version.

Here we give the definitions of query-based and code-based API recommendation:

Query-based API recommendation: Approaches for query-based API recommendation aim at providing related APIs to developers given a query which describes programming requirements in natural language. The approaches can inform developers which API to use for a programming task.

Code-based API recommendation: Approaches for code-based API recommendation aim at predicting the next API given the code surrounding the point of prediction. They can directly improve the efficiency of coding.

APIBench-Q is located in the folder APIBench/APIBench_Q. There are a Python folder and a Java folder in it for the Python and Java version of APIBench-Q.

Currently there are two json files in each version:

Original{Java,Python}Queries.json: It contains the original queries, corrsponding APIs and API classes, along with the source the query collected from (Stack Overflow or Tutorial Websites).

Reformulated{Java,Python}Queries.json : It contains all the elements in Original{Java,Python}Queries.json. Besides, it contains the reformulated queries derive from the original queries. The reformulated queries are wrapped by @.

Currently we apply the following query reformulation techniques to process the original queries:

Number of queries in APIBench-Q:

APIBench-C is located in the folder APIBench/APIBench_C. There are a Python folder and a Java folder in it for the Python and Java version of APIBench-C.

Currently there are two zip files and a json file in each version:

{Java,Python}_MetaData.zip: It contains the metadata json files for code in each domain. The long, normal and short kewords indicate the function with extremely long, moderate and extremely short lengths.

{Java, Python}_Code.zip: It contains the source code of repositories we mined from GitHub at April, 2021. Only .py and .java files are reserved.

{Java, Python}RepoInfo.json: It contains the information of Github repositories we collected, including the lines of code, code ratio, domain, files, forks, stars, addresses, etc.

The structure of metadata:

"$filename$": {
		"$classname$" or "@Global@": {
      "$funcname$":{
        "@FuncLoc@":[
          $startlineno$,
          $endlineno$
        ],
        "$APIname$":[
          [
            $lineno$,
            $columnno$,
            "pure" or "attr",  #whether called from a class
            "Front", "Mdiddle" or "Back" #location of recommendation point
          ],
          ...
          ,
          "Standard", "User-defined", "Popular", or "Unknown" #category of APIs
        ]
      }
    }
}

Statistics of APIBench-C:

To facilitate further research on API recommendation and reduce the burden of re-implementing different baselines, we release all the evaluation results and outputs of 11 baselines along with 4 IDEs discussed in the paper.

Please go to the experiment_resultsfolder for further detailed information.

If you use our benchmark dataset and related experiment results or code, please cite us:

@misc{peng2021revisiting,
      title={Revisiting, Benchmarking and Exploring API Recommendation: How Far Are We?}, 
      author={Yun Peng and Shuqing Li and Wenwei Gu and Yichen Li and Wenxuan Wang and Cuiyun Gao and Michael Lyu},
      year={2021},
      eprint={2112.12653},
      archivePrefix={arXiv},
      primaryClass={cs.SE}
}

If you have any questions, please contact [email protected].