README.md

metadata

dataset_info:
  features:
    - name: commit
      dtype: string
    - name: old_file
      dtype: string
    - name: new_file
      dtype: string
    - name: old_contents
      dtype: string
    - name: new_contents
      dtype: string
    - name: subject
      dtype: string
    - name: message
      dtype: string
    - name: lang
      dtype: string
    - name: license
      dtype: string
    - name: repos
      dtype: string
    - name: ndiff
      dtype: string
    - name: instruction
      dtype: string
    - name: content
      dtype: string
  splits:
    - name: train
      num_bytes: 113752028
      num_examples: 22602
  download_size: 48124127
  dataset_size: 113752028
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - text-generation
tags:
  - code
license: mit
pretty_name: CanItEdit
language:
  - code

EditPackFT

EditPackFT is a dataset built for training LLMs on the task of instructional code editing. The mail columns are:

1. old_contents the code before the edit
2. instruction the instruction to transform the before code into the after code
3. new_contents the code after the edit
4. content a pre-formatted training window that can be used to train an LLM with prompts in the format of: <before><instruction><after>

This dataset has been filtered from CommitPackFT. For more detail, see our paper, and our GitHub repository.

Citation

If you use our work, please cite our paper as such:

@misc{cassano2023edit,
      title={Can It Edit? Evaluating the Ability of Large Language Models to Follow Code Editing Instructions}, 
      author={Federico Cassano and Luisa Li and Akul Sethi and Noah Shinn and Abby Brennan-Jones and Anton Lozhkov and Carolyn Jane Anderson and Arjun Guha},
      year={2023},
      eprint={2312.12450},
      archivePrefix={arXiv},
      primaryClass={cs.SE}
}