CarperAI / Code-Pile

This repository contains all the code for collecting large scale amounts of code from GitHub.
MIT License
105 stars 29 forks source link
dataset

Code-Pile

pytest

This repository contains the processing scripts to scrape/process the code-pile dataset.

Table of Contents

Project Description

Check out The code pile proposal

The Code-Pile will be released similar to "the pile" as a folder of .jsonl.zst files, see lm-dataformat

How to use the Code-Pile

It's not finished, ask on discord

How to Contribute

Think about the most usefull Code-data for the next generation of textual Code Models.

The most valuable dataset properties (use your own judgment) are:

  1. Open License
  2. Data quality
  3. Dataset size
  4. Data variance/variety/nicheness
  5. Ease of obtaining/processing

To add a new dataset, open a Issue under given dataset-request template. Gather all the related informations appropriate to it. Use the issue to track.

Check if there is existing Code or someone already working on it: See Additional Resources

  1. Eleuthers Pile V1 Repos
  2. Ask on Carper #code-pile
  3. Ask on Eleuther
  4. Consult the linked Spreadsheets below

Then implement it through the following steps:

  1. Fork this repo
  2. Use the working branch
  3. Read the shared classes in datasets.py and codepile.py
  4. Create mvp/example for your dataset
  5. Create a pull request
  6. Keep building the data-domain specific classes and repeat

Citation Placeholder:

@misc{Code-Pile,
  author = {},
  doi = {},
  month = {},
  title = {},
  url = {https://github.com/CarperAI/Code-Pile},
  version = {},
  year = {2022}
}

Additional Resources

Closely related projects:

Previous work: