AdityaNarayan/HS-Repo-Curriculum-Learning

Hyperswitch Curriculum Learning Dataset (Unbroken)

A comprehensive dataset for continued pre-training (CPT) of large language models on the Hyperswitch payment processing codebase, organized into curriculum learning phases with complete, unbroken entries.

🎯 Dataset Overview

This dataset contains the complete Hyperswitch repository knowledge extracted from:

• Source code files (.rs, .toml, .yaml, .json, .md)
• Git commit history with full diffs
• GitHub Pull Requests with reviews and discussions
• Test-implementation pairs

Key Feature: Unlike the chunked version, each entry is stored complete without breaking at token boundaries, allowing dynamic chunking during training for any sequence length (8K, 16K, 32K, 64K+).

📊 Dataset Structure

Curriculum Learning Phases

The dataset is organized into 3 progressive phases:

Phase 1: Code Foundation (phase1_foundation.jsonl)

• Content: Repository files + test-implementation pairs
• Purpose: Learn codebase structure, syntax, and testing patterns
• Training: 2 epochs
• Entries: Complete files and test pairs (unbroken)

Phase 2: Evolution Patterns (phase2_evolution.jsonl)

• Content: Git commits (chronological) + small PRs
• Purpose: Understand code evolution, change patterns, and incremental development
• Training: 2-3 epochs
• Entries: Complete commits with full diffs, small PRs (unbroken)

Phase 3: PR Mastery (phase3_pr_mastery.jsonl)

• Content: Medium and large PRs with reviews and discussions
• Purpose: Master complex changes, code review practices, and collaboration patterns
• Training: 3-4 epochs
• Entries: Complete PRs with all reviews and comments (unbroken)

📁 Data Format

Each entry is a single JSON object per line (JSONL format):

File Entry

{
  "type": "file",
  "path": "crates/hyperswitch_connectors/src/connectors/paypal/transformers.rs",
  "size_bytes": 140434,
  "training_content": "// File: crates/hyperswitch_connectors/src/connectors/paypal/transformers.rs\n\n<complete_file_content>"
}

Commit Entry

{
  "type": "commit",
  "commit_hash": "73203ebd05beab57f243e8460f259707bb856921",
  "author": "vasanthp-jus",
  "date": "2025-11-27T12:18:26+05:30",
  "message": "fix-postman-collection",
  "training_content": "Commit: \"fix-postman-collection\"\nAuthor: vasanthp-jus\nDate: 2025-11-27T12:18:26+05:30\n\nDiff:\n<complete_git_diff>"
}

PR Entry

{
  "type": "pr_diff",
  "pr_number": 1234,
  "title": "Add PayPal connector support",
  "state": "merged",
  "author": "developer-name",
  "created_at": "2025-11-15T10:30:00Z",
  "training_content": "PR #1234: Add PayPal connector support\n\n<description>\n\nReviews:\n<complete_reviews>\n\nComments:\n<complete_comments>"
}

Test Pair Entry

{
  "type": "test_pair",
  "test_file": "crates/router/tests/connector_tests.rs",
  "impl_file": "crates/router/src/connector.rs",
  "training_content": "Test-Implementation Pair:\n\nTest: <test_content>\n\nImplementation: <impl_content>"
}

🔢 Dataset Statistics

Phase	Entries	Content Types	Avg Entry Size
Phase 1	~15K	Files, Test Pairs	Varies (complete files)
Phase 2	~5K	Commits, Small PRs	Varies (complete commits/PRs)
Phase 3	~1K	Medium/Large PRs	Large (complete PR threads)

Total: ~21K complete, unbroken entries

💡 Unbroken vs Chunked

Unbroken (This Dataset)

✅ Complete semantic units preserved
✅ No artificial breaks in code/diffs
✅ Flexible for any sequence length
✅ Chunk dynamically during training
✅ Smaller dataset file size (no overlap)

Chunked (Alternative)

• Pre-chunked at fixed token limit (e.g., 8K)
• Ready for immediate training
• Fixed sequence length
• Includes chunk overlap for continuity

🚀 Usage

Loading the Dataset

import json

def load_phase(phase_file):
    """Load a curriculum phase."""
    entries = []
    with open(phase_file, 'r', encoding='utf-8') as f:
        for line in f:
            entries.append(json.loads(line))
    return entries

# Load Phase 1
phase1 = load_phase('phase1_foundation.jsonl')

Dynamic Chunking for Training

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("your-model")
max_length = 32768  # 32K tokens

def chunk_entry(entry, tokenizer, max_length):
    """Chunk a complete entry for training."""
    text = entry['training_content']
    
    # Tokenize
    tokens = tokenizer(text, truncation=False, return_tensors='pt')
    
    # Split into chunks if needed
    chunks = []
    token_ids = tokens['input_ids'][0]
    
    for i in range(0, len(token_ids), max_length):
        chunk = token_ids[i:i + max_length]
        chunks.append(chunk)
    
    return chunks

# Process entries
for entry in phase1:
    chunks = chunk_entry(entry, tokenizer, max_length)
    for chunk in chunks:
        # Use chunk for training
        pass

Recommended Training Schedule

# Phase 1: Code Foundation (2 epochs)
train(phase1_foundation, epochs=2, lr=1e-5)

# Phase 2: Evolution Patterns (2-3 epochs)
train(phase2_evolution, epochs=3, lr=8e-6)

# Phase 3: PR Mastery (3-4 epochs)
train(phase3_pr_mastery, epochs=4, lr=5e-6)

🎓 Curriculum Learning Benefits

• Progressive complexity: Start simple, increase difficulty
• Better convergence: 25-40% improvement over random training
• Domain adaptation: Learn repository-specific patterns
• Code understanding: Syntax → Changes → Collaboration
• Efficient training: Focused learning objectives per phase

📝 Technical Details

Repository

• Source: Hyperswitch
• Language: Primarily Rust
• Domain: Payment processing, financial technology
• Components: Connectors, API models, routing logic, state machines

Data Collection

• Files: Pattern-based extraction (Rust, TOML, YAML, JSON, Markdown)
• Commits: Full git history from repository inception
• PRs: Merged and closed PRs with reviews and comments via GitHub API
• Tests: Automatic pairing of test files with implementations

🔧 Sequence Length Flexibility

This unbroken dataset works with any sequence length:

Sequence Length	Use Case	Chunking Strategy
8K tokens	Base models	Chunk with overlap
16K tokens	Extended context	Fewer chunks needed
32K tokens	Long context models	Most files fit whole
64K+ tokens	Ultra-long context	Complete commits/PRs

🙏 Acknowledgments

• Hyperswitch Team at Juspay for the amazing open-source payment processing platform
• Dataset curated and organized by Aditya Narayan
• Dataset generated using custom extraction pipeline with curriculum organization

📧 Contact & Citation

If you use this dataset, please cite:

@dataset{hyperswitch_curriculum2025,
  title = {AdityaNarayan/HS-Repo-Curriculum-Learning},
  author = {Aditya Narayan},
  year = {2025},
  url = {https://huggingface.co/datasets/AdityaNarayan/HS-Repo-Curriculum-Learning},
  publisher = {HuggingFace},
  note = {Dataset derived from Hyperswitch repository}
}