msme-document-presence-xgboost

MSME Document Presence Detection Model

Overview

This repository contains production-grade XGBoost models designed to detect the presence of mandatory documents in MSME arbitration cases based on OCR-extracted text.

The system performs binary classification for the following documents:

• Invoice
• Purchase Order
• Delivery Proof
• GST Certificate
• Contract

Each document is modeled independently as a separate classifier.

---

Model Architecture

• Algorithm: XGBoost (Gradient Boosted Trees)
• Feature Extraction: TF-IDF (1–2 n-grams)
• Max Features per model: 3000
• Independent model per document type
• Stratified train-test split
• Hard negative augmentation included
• Severe OCR corruption simulation included

---

Training Data

The model was trained on a synthetic and augmented dataset consisting of:

• 5,000 LLM-generated structured OCR samples
• OCR distortion simulation
• Keyword masking
• Partial truncation
• Cross-document contamination
• Line shuffling
• Hard negative construction
• Class imbalance simulation

Final training dataset size: approximately 10,000 samples.

---

Performance

Average performance across all document classifiers:

• Precision: 0.992
• Recall: 0.987
• F1 Score: 0.990
• ROC-AUC: 0.999
• False Negative Rate: < 2%

Performance evaluated using stratified 80/20 split.

---

Inference

Each model expects OCR-extracted raw text for a specific document type.

Output per document:

• Binary prediction (0 = Missing, 1 = Present)
• Probability score
• Optional SHAP-based explainability (external implementation)

Completeness Score can be computed as:

completeness = (documents_present / required_documents) × 100

---

Intended Use

This model is suitable for:

• MSME arbitration automation
• Legal document validation pipelines
• OCR post-processing systems
• Document completeness scoring engines
• Hybrid rule + ML legal systems

---

Limitations

• Trained primarily on synthetic and augmented OCR data
• Real-world scanned PDFs may introduce unseen distortions
• Extreme low-quality scans may reduce recall
• Contract optionality logic must be implemented externally
• Not intended for semantic contract analysis

---

Ethical Considerations

The model was trained exclusively on synthetic data. No real personal, financial, or legal records were used.

---

Future Work

• Fine-tuning on real arbitration case documents
• Probability calibration
• Threshold optimization per document type
• Model drift monitoring
• Ensemble rule + ML integration
• ONNX export for optimized inference

---

License

This project is released under the MIT License.