porosdata-processor 0.2.3

Academic document intelligent cleaning pipeline for AI for Science, ensuring MinerU parsed data meets LLM input standards

PorosData-Processor

Academic Literature Data Engineering Toolkit - Specializes in cleaning MinerU outputs (scientific literature) and evaluating Token efficiency for LLMs, supporting the complete "AI for Science" workflow.

🎯 Project Positioning

In the AI for Science field, high-quality academic data preprocessing is the foundation for models to understand scientific literature. PorosData-Processor serves as a data engineering tool that specifically addresses the "last mile" problem from MinerU document parsing to LLM input, ensuring academic documents achieve maximum Token efficiency while maintaining integrity.

🌟 Core Features

• 🔬 Professional Academic Cleaning: Intelligently handles LaTeX formulas, control characters, citation formats, and other academic document-specific issues
• ⚡ Multi-process Parallel Processing: Cross-platform concurrent processing based on pathlib and spawn methods, supporting Windows/Linux
• 📊 Real-time Token Evaluation: Integrated GPT-2 tokenizer, providing precise Token compression rate calculations
• 🛡️ Intelligent Protection Mechanism: 20% compression rate threshold protection, ensuring text semantic integrity
• 🔄 Self-healing Quality Assurance: Automatic detection and repair of data corruption during processing

🚀 Quick Start

Environment Requirements

pip install porosdata-processor ijson tiktoken psutil

Process MinerU Data

# Process all JSON files in the data/mineru_output_raw_data directory
# Automatically configure HF_ENDPOINT mirror for Chinese users to accelerate downloads
python run_processor.py --enable-evaluation

Output Data Format

Processed JSON files contain standardized fields:

• text: Cleaned academic text, suitable for LLM input
• original_text: Original input text, for quality comparison
• healed_count: Self-healing repair count, reflecting data quality

📈 Performance Showcase

Materials Science Data Pilot Test Results

Test Metric	Value	Description
Files Processed	3 files	MinerU-parsed academic paper JSON files
Items Processed	127 items	Including text, formulas, tables, and other structured content
Avg Token Compression Rate	0.098	90.2% Token Savings
Processing Time	0.456s	Multi-process parallel processing efficiency
Memory Peak	204.2MB	Streaming processing ensures memory efficiency

Key Insights:

• Token compression rate reaches 0.098 (90.2% reduction), significantly reducing LLM inference costs
• Processing speed of 278.5 items/second, suitable for large-scale academic data processing
• Memory peak usage of only 204.2MB, supporting TB-level data processing

🛠️ Academic Tools Suite

The project includes a complete academic data processing toolchain:

Core Processing Scripts

• Batch Processing: academic_tools/standalone/batch_process.py
• Single File Processing: academic_tools/standalone/process_single_json.py
• Compatibility Processing: academic_tools/standalone/process_with_cleanlit.py

Advanced Configuration Options

# Custom input/output directories
python run_processor.py --input-dir ./data/input --output-dir ./data/output

# Basic cleaning only (no Token evaluation)
python run_processor.py

# Force reprocessing of all files
python run_processor.py --enable-evaluation --force-reprocess

📚 Technical Documentation

• Architecture Design - Core components and implementation principles
• Usage Guide - Detailed API and configuration instructions
• Testing Guide - Development and testing environment setup

🔬 Technical Specifications

Supported Data Formats

• Input: MinerU-parsed JSON format academic documents
• Output: Standardized JSON, compatible with LLM training and inference
• Encoding: UTF-8 cross-platform support, automatic control character handling

Quality Assurance Mechanisms

• Compression Rate Protection: ≤20% threshold ensures text semantic integrity
• Boundary Self-healing: Automatic detection and repair of Shield protection anomalies
• Integrity Auditing: Multi-layer verification ensures data quality

📄 Open Source License

MIT License - see LICENSE file for details

📖 Citation

If you use PorosData-Processor in your research, please cite:

@software{porosdata_processor,
  title = {PorosData-Processor: Academic Document Intelligent Cleaning Pipeline},
  author = {YE, Kivent},
  year = {2025},
  url = {https://github.com/KiventYip/PorosData-doc},
  version = {0.2.2}
}

🤝 Contributions and Feedback

Issues and Pull Requests are welcome! The project adopts a data-driven development philosophy, and any performance optimization suggestions will be seriously considered.