ethancaballero / description2code

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Description2Code dataset and (very messy) scrapers

Description2Code is dataset of ~7764 programming challenges scraped from CodeChef, Codeforces, & Hackerearth in 2016. Each programing challenge in this dataset provides a problem description and multiple solutions and multiple test cases.

Link to download dataset: https://www.dropbox.com/s/zwj6u4caehf54s0/description2code_current.zip

Story of dataset creation: openai/requests-for-research#5

Very messy code that was used for scraping Codechef, Codeforces, Hackerearth, & Topcoder in 2016 is located in scrapers folder of this repo. I don't remember why Topcoder isn't included in the dataset.

License

I don't "own" the data scraped. CodeChef and CodeForces and HackerEarth or their users or something technically own it (or maybe they don't; idk). I don't know what the legal_rights/data_licenses of CodeChef and CodeForces and HackerEarth and their users are. I'm fine with you using this dataset and scraping code however you want.

Citation

If you find this dataset (or scrapers) useful, please cite us in your work ; BibTex to use:

@misc{Caballero_Description2Code_Dataset_2016,
author = {Caballero, Ethan and OpenAI, . and Sutskever, Ilya},
doi = {10.5281/zenodo.5665051},
month = {8},
title = {{Description2Code Dataset}},
url = {https://github.com/ethancaballero/description2code},
year = {2016}
}

Notes about dataset that download link contains

CodeChef

  • Contains curriculum of problems increasing as easy < medium < hard < harder < hardest
  • External folder contains 2070 problems of varying difficulty
  • does not contain samples and annotations for now
  • 3201 total

CodeForces

  • Problem's difficulty relative to other problems in individual competition is indicated by suffix letter on end of folder ( A < B < C < D < E < F < G < I < J ); individual competition is signified by prefix number at start of folder
  • '_tags.txt' contains list of which types of algorithmic techniques are needed to solve each problem; could be used to benchmark which types of problems your model is/isn't capable of solving or possibly as a supervision signal.
  • 2128 total

HackerEarth

  • descriptions in 'problem_normal' are written slightly better than in 'problem_college'.
  • '_tags.txt' contains list of the difficulty of and which types of algorithmic techniques are needed to solve each problem; could be used to benchmark which types of problems your model is/isn't capable of solving or possibly as a supervision signal.
  • 2435 total

General Usage

  • description folder contains input for model; solutions_* (c++ or python depending on what you want to train on) contains target output for model.
  • Use Example Input(s)/Output(s) near bottom of 'description.txt' to test whether generated code is correct.
  • Train on multiple solutions provided for each problem to help your model generalize.
  • 'Samples' folders contain several input/output samples that could be used to reward model when it generates code that correctly processes sample input into sample output. Several input/output examples are contained in each samples folder to prevent model from overfitting to generated code that could only correctly process a single sample input to sample output.
  • 'description_annotated.txt' is version of description that annotates tokens (with dagger (U+2020) and double_dagger (U+2021) at beginning and end of token(s) respectively) that are meant to be program variables.
  • Use several sample input/output from samples folder to reward model for generating code that passes all possible test cases.
  • versions of collected solution source codes are Python 2 and C++ 4.3.2
  • 7764 total

About


Languages

Language:Python 100.0%