Dataset and Code for: Code problem similarity detection using code clones and pretrained models

This dataset complements the following study: Code problem similarity detection using code clones and pretrained models (SCSE22-0384). This study explores a new approach of detecting similar algorithmic-style code problems from websites such as LeetCode and Codeforces, by comparing the similarity of the solution source codes, an application of type IV code clone detection. It is based on 107,000 submissions in 3 different languages (Python, C++ and Java) from 3,000 problems on Codeforces between 2020 to 2023. Experiments were carried out using 3 different pre-trained models on this dataset (C4-CodeBERT, GraphCodeBERT, UniXcoder). UniXcoder performed the best with an F1 score of 0.905. As such, UniXcoder was used as the backbone of the code problem similarity checker (CPSC) which is used to identify the top similar problems (out of all the problems in the dataset) to an input source code. Based on the tests conducted in this project, his approach achieves state-of-the-art results when it comes to detecting similarity between various code problems. More research can be done, in domains where type IV code clone detection can be useful.