Production-Grade NER Model

Single Entry Point Design - One command for everything: train, test, and deploy!

🚀 Quick Start

# Full pipeline: Train → Test → Interactive
python main.py --full-pipeline --real-world

# Individual modes
python main.py --real-world          # Train only (default)
python main.py --test               # Test trained model
python main.py --interactive        # Interactive terminal

📋 Commands Overview

Command	Description
python main.py --real-world	Train robust model with overfitting fixes
python main.py --test	Run comprehensive testing suite
python main.py --interactive	Start interactive NER interface
python main.py --full-pipeline --real-world	Complete workflow

🛠️ Project Structure

project_b/
├── main.py                          # 🎯 SINGLE ENTRY POINT
├── tools/                           # Testing & interaction utilities
│   ├── test_model_comprehensive.py  # Comprehensive testing suite
│   └── interactive_ner.py           # Interactive terminal interface
├── model.py                         # Model architecture
├── train.py                         # Training utilities
├── utils.py                         # Data processing
├── evaluate.py                      # Evaluation metrics
└── advanced_real_world_training.py  # Overfitting fixes

🎯 Training Options

Real-World Training (Recommended)

python main.py --real-world --epochs 8 --batch-size 8 --learning-rate 1.5e-5

Fixes Applied:

• ✅ Surface pattern corruption (breaks @ .com memorization)
• ✅ Realistic targets: EMAIL 85-90%, PHONE 82-88%, PERSON 88-94%
• ✅ Stronger regularization (weight_decay 0.02, label_smoothing 0.1)
• ✅ More frequent evaluation (every 100 steps)

Standard Training

python main.py                      # Basic training
python main.py --epochs 4           # Custom epochs
python main.py --batch-size 16      # Custom batch size

🧪 Testing

Comprehensive Testing Suite

python main.py --test

Tests Include:

• ✅ Basic entity recognition (PER, ORG, LOC, MISC)
• ✅ EMAIL/PHONE detection accuracy
• ✅ Overfitting detection (corrupted patterns)
• ✅ Edge cases (empty text, long sequences, special chars)
• ✅ Performance analysis with 80% success threshold

Output: model_test_report.json with detailed results

🚀 Interactive Mode

python main.py --interactive

Features:

• 🎨 Colored entity highlighting
• ⚡ Real-time processing with timing
• 📊 Built-in commands: help, examples, stats, test
• 🛡️ Safety checks (requires testing first)

🔄 Full Pipeline

python main.py --full-pipeline --real-world

Workflow:

1. Training - Robust real-world training with overfitting fixes
2. Testing - Comprehensive validation (must pass ≥80%)
3. Interactive - Deploy only if testing passes

📊 Expected Performance

Entity Type	Target F1	Note
EMAIL	85-90%	Not 99%+ (overfitting)
PHONE	82-88%	Not 99%+ (overfitting)
PERSON	88-94%	Not 99%+ (overfitting)
OVERALL	86-92%	Not 98%+ (overfitting)

🔧 Advanced Options

# Custom model
python main.py --model roberta-base --real-world

# More synthetic data
python main.py --synthetic-count 10000 --real-world

# Quick test run
python main.py --epochs 2 --batch-size 4

📁 File Organization

• main.py - Central orchestrator, single entry point
• tools/ - Separate but integrated utilities
• production-ner-model-final/ - Trained model output
• checkpoints/ - Training checkpoints
• mlruns/ - MLflow experiment tracking

🎯 Production Deployment

1. Train: python main.py --real-world
2. Validate: python main.py --test (≥80% required)
3. Deploy: python main.py --interactive

Safety: Interactive mode requires testing completion first!