benchmarks/README.md

CodeReality-1T Evaluation Benchmarks

This directory contains demonstration benchmark scripts for evaluating models on the CodeReality-1T dataset.

Available Benchmarks

1. License Detection Benchmark

File: license_detection_benchmark.py

Purpose: Evaluates automated license classification systems on deliberately noisy data.

Features:

• Rule-based feature extraction from repository content
• Simple classification model for demonstration
• Performance metrics on license detection accuracy
• Analysis of license distribution patterns

Usage:

cd /path/to/codereality-1t/eval/benchmarks
python3 license_detection_benchmark.py

Expected Results:

• Low accuracy due to deliberately noisy dataset (0% license detection by design)
• Demonstrates robustness testing for license detection systems
• Outputs detailed distribution analysis

2. Code Completion Benchmark

File: code_completion_benchmark.py

Purpose: Evaluates code completion models using Pass@k metrics on real-world noisy code.

Features:

• Function extraction from Python, JavaScript, Java files
• Simple rule-based completion model for demonstration
• Pass@1, Pass@3, Pass@5 metric calculation
• Multi-language support with language-specific patterns

Usage:

cd /path/to/codereality-1t/eval/benchmarks
python3 code_completion_benchmark.py

Expected Results:

• Baseline performance metrics for comparison
• Language distribution analysis
• Quality scoring of completions

Benchmark Characteristics

Dataset Integration

• Data Source: Loads from /mnt/z/CodeReality_Final/unified_dataset by default
• Sampling: Uses random sampling for performance (configurable)
• Formats: Handles JSONL repository format from CodeReality-1T

Evaluation Philosophy

• Deliberately Noisy: Tests model robustness on real-world messy data
• Baseline Metrics: Provides simple baselines for comparison (not production-ready)
• Reproducible: Deterministic evaluation with random seed control
• Research Focus: Results show challenges of noisy data, not competitive benchmarks

Extensibility

• Modular Design: Easy to extend with new benchmarks
• Configurable: Sample sizes and evaluation criteria can be adjusted
• Multiple Languages: Framework supports cross-language evaluation

Configuration

Data Path Configuration

Update the data_dir variable in each script to point to your CodeReality-1T dataset:

data_dir = "/path/to/your/codereality-1t/unified_dataset"

Sample Size Adjustment

Modify sample sizes for performance tuning:

sample_size = 500  # Adjust based on computational resources

Output Files

Each benchmark generates JSON results files:

• license_detection_results.json
• code_completion_results.json

These contain detailed metrics and can be used for comparative analysis.

Sample Results

Example results are available in ../results/:

• license_detection_sample_results.json - Baseline license detection performance
• code_completion_sample_results.json - Baseline code completion metrics

These demonstrate expected performance on CodeReality-1T's deliberately noisy data.

Requirements

Python Dependencies

pip install json os re random typing collections

System Requirements

• Memory: Minimum 4GB RAM for default sample sizes
• Storage: Access to CodeReality-1T dataset (3TB)
• Compute: Single-core sufficient for demonstration scripts

Extending the Benchmarks

Adding New Tasks

1. Create new Python file following naming convention: {task}_benchmark.py
2. Implement standard evaluation interface:

   def load_dataset_sample(data_dir, sample_size)
   def run_benchmark(repositories)
   def print_benchmark_results(results)
   

3. Add task-specific evaluation metrics

Supported Tasks

Current benchmarks cover:

• License Detection: Classification and compliance
• Code Completion: Generation and functional correctness

Framework Scaffolds (PLANNED - Implementation Needed):

• bug_detection_benchmark.py - Bug detection on commit pairs (scaffold only)
• cross_language_translation_benchmark.py - Code translation across languages (scaffold only)

Future Planned Benchmarks - Roadmap:

• v1.1.0 (Q1 2025): Complete bug detection and cross-language translation implementations
• v1.2.0 (Q2 2025): Repository classification and domain detection benchmarks
• v1.3.0 (Q3 2025): Build system analysis and validation frameworks
• v2.0.0 (Q4 2025): Commit message generation and issue-to-code alignment benchmarks

Community Priority: Framework scaffolds ready for community implementation!

Performance Notes

Computational Complexity

• License Detection: O(n) where n = repository count
• Code Completion: O(n*m) where m = average functions per repository

Optimization Tips

1. Sampling: Reduce sample_size for faster execution
2. Filtering: Pre-filter repositories by criteria
3. Parallelization: Use multiprocessing for large-scale evaluation
4. Caching: Cache extracted features for repeated runs

Research Applications

Model Development

• Robustness Testing: Test models on noisy, real-world data
• Baseline Comparison: Compare against simple rule-based systems
• Cross-domain Evaluation: Test generalization across domains

Data Science Research

• Curation Methods: Develop better filtering techniques
• Quality Metrics: Research automated quality assessment
• Bias Analysis: Study representation bias in large datasets

Citation

When using these benchmarks in research, please cite the CodeReality-1T dataset:

@misc{codereality2025,
  title={CodeReality-1T: A Large-Scale Deliberately Noisy Dataset for Robust Code Understanding},
  author={Vincenzo Gallo},
  year={2025},
  publisher={Hugging Face},
  howpublished={\\url{https://huggingface.co/vinsblack}},
  note={Version 1.0.0}
}

Support

• Issues: https://github.com/vinsguru/codereality-1t/issues
• Contact: [email protected]
• Documentation: See main dataset README and documentation