kotlarmilos/repository-learning

Repository Learning Training Dataset

This dataset contains training data extracted from GitHub repositories for training context-aware code review models. The dataset supports three primary machine learning tasks: contrastive learning, fine-tuning, and semantic indexing.

Dataset Overview

Purpose: Enable training of AI models that understand repository-specific code review patterns and provide contextual feedback.

Source: GitHub repositories with rich pull request history and review comments.

Dataset Structure

{repository-name}/
├── contrastive/
│   ├── changed_files_001.json      # Files changed together (positive pairs)
│   ├── changed_files_002.json   
│   └── ...
├── fine_tune/
│   ├── pr_reviews_001.jsonl        # Instruction-following format
│   ├── pr_reviews_002.jsonl
│   └── ...
├── index/
│   ├── functions.json              # AST-extracted function metadata
└── manifest.json                   # Processing metadata

Data Components

1. Contrastive Learning Data (/contrastive/)

Format: JSON files containing file groupings for contrastive learning.

Purpose: Learn semantic relationships between code files based on change patterns.

Structure:

{
  "pr_12345": [
    "src/components/Button.tsx",
    "src/styles/button.css", 
    "tests/Button.test.tsx"
  ],
  "pr_12346": [
    "src/api/user.py",
    "src/models/user.py",
    "tests/test_user.py"
  ]
}

Usage: Files changed together form positive pairs; files from different PRs form negative pairs for contrastive learning.

2. Fine-Tuning Data (/fine_tune/)

Format: JSONL files with instruction-following examples.

Purpose: Adapt language models to repository-specific review patterns and conventions.

Structure:

{
  "prompt": "Code diff:\n```diff\n+def calculate_score(user_data):\n+    return sum(user_data.values())\n```\nPrevious comments:\n- alice: Consider input validation\n\nPlease write a code review comment:",
  "completion": "Good addition! I'd suggest adding type hints and handling edge cases where user_data might be empty or contain non-numeric values."
}

Features:

• Chronological conversation context
• Multi-turn review discussions
• Repository-specific terminology and patterns
• Code diff context with surrounding discussion

3. Semantic Index Data (/index/)

Format: JSON metadata with function definitions and embeddings.

Purpose: Enable fast semantic search across repository functions and documentation.

Structure (functions.json):

[
  {
    "file": "src/utils/parser.py",
    "name": "parse_diff_hunk", 
    "start_line": 45,
    "end_line": 67,
    "code": "def parse_diff_hunk(hunk_text: str) -> DiffHunk:\n    # Function implementation...",
  }
]

Components:

• AST Extraction: Tree-sitter parsers for different programming languages

Data Generation Pipeline

Data Statistics

Repository	PRs	Review Comments	Functions	Languages
dotnet/xharness	100	50	1500	C#
dotnet/runtime	N/A	N/A	N/A	C#, c, c++

Usage Examples

If you use this dataset, please refer to https://github.com/kotlarmilos/repository-learning