README.md

metadata

dataset_info:
  features:
    - name: repo
      dtype: string
    - name: commit
      dtype: string
    - name: path
      dtype: string
    - name: lang
      dtype: string
    - name: license
      dtype: string
    - name: message
      dtype: string
    - name: old_code
      dtype: string
    - name: new_code
      dtype: string
    - name: diff_string
      dtype: string
    - name: n_added
      dtype: int64
    - name: n_removed
      dtype: int64
    - name: n_hunks
      dtype: int64
    - name: change_kind
      dtype: string
  splits:
    - name: test
      num_bytes: 3425951
      num_examples: 1000
  download_size: 1664793
  dataset_size: 3425951
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

diff-xyz-noisy

A noisy version of diff-xyz dataset. This dataset is obtained from the original one by replacing alpha symbols with another alpha symbol using a random character mapping. Components: the original file contents (old_code), the modified contents (new_code), and the unified diff between them (diff_string).

This format enables evaluation of LLM capabilities on three code editing tasks:

• Apply: Given old code and diff, generate new code
• Anti-Apply: Given new code and diff, generate old code
• Diff-Generation: Given old and new code, generate the unified diff