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: n_added
      dtype: int64
    - name: n_removed
      dtype: int64
    - name: n_hunks
      dtype: int64
    - name: change_kind
      dtype: string
    - name: udiff
      dtype: string
    - name: udiff-h
      dtype: string
    - name: udiff-l
      dtype: string
    - name: search-replace
      dtype: string
  splits:
    - name: test
      num_bytes: 5891980
      num_examples: 1000
  download_size: 2695508
  dataset_size: 5891980
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

Diff-XYZ

This is a dataset for the paper: Diff-XYZ: A Benchmark for Evaluating Diff Understanding.

Diff-XYZ contains 1,000 real-world code edits sampled and filtered from the CommitPackFT dataset.
Each example provides three components: the original file contents (old_code), the modified contents (new_code), and multiple diff representations (udiff, udiff-h, udiff-l, and search-replace).

These formats enable 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 diff

How it's built

Examples were systematically filtered from CommitPackFT to ensure high quality:

• Scope: Single-file changes only, excluding binary files, vendor directories, and generated code
• Quality: Removed trivial changes (whitespace-only) and likely test files
• Size: Required 40+ lines in at least one version of the code (old or new); excluded very large files (1000+ lines)
• Sampling:
  • Target 50/50 split between single-hunk and multi-hunk edits (hunks counted via @@ markers in the unified diff).
  • Within each hunk group, stratify by change size (lines added + removed) using the 40th/80th percentiles of that group, targeting ≈40% small, 40% medium, 20% large.
  • Cap examples to ≤5 per repository per language.

Dataset Statistics

• Total examples: 1,000
• Languages: Python (200), JavaScript (200), Java (200), Kotlin (200), Rust (200)

Hunks:

• 1 hunk: 500 (50.0%)
• 2+ hunks: 500 (50.0%)

Change size (added + removed), stratified within hunk groups:

• Small (≤40th pct): 424 (42.4%)
• Medium (40th–80th): 388 (38.8%)
• Large (>80th): 188 (18.8%)

Change type:

• Mixed (additions + deletions): 815 (81.5%)
• Add-only: 163 (16.3%)
• Delete-only: 22 (2.2%)

Repository diversity: 891 unique repositories.

Schema

Field	Type	Description
repo	string	Repository identifier (e.g., "owner/name")
commit	string	Commit SHA that produced the change
path	string	File path relative to repo root
lang	string	One of: python, javascript, java, kotlin, rust
license	string	SPDX license identifier of the repo
message	string	Original commit message
old_code	string	Complete file contents before change
new_code	string	Complete file contents after change
n_added	int	# of + lines (excluding headers)
n_removed	int	# of - lines (excluding headers)
n_hunks	int	# of @@ (hunk) sections
change_kind	string	add_only, del_only, or mixed
udiff	string	Standard unified diff (1-line context, numeric hunk headers)
udiff-h	string	Unified diff with relaxed hunk headers written as @@ ... @@
udiff-l	string	Unified diff with explicit line markers: ADD, DEL, and CON
search-replace	string	Search/replace representation: pairs of SEARCH/REPLACE edit blocks