Phi-4 Information Extraction Demo

Interactive CLI demo for extracting entities, relations, and knowledge graph triples from business text using a fine-tuned Phi-4 model (15B parameters).

Features

• Named Entity Recognition (NER): Extract 40+ business entity types (companies, products, people, events, financial metrics, etc.)
• Relation Extraction (RE): Identify relationships between entities
• Triple Generation: Output structured knowledge graph triples
• Completely Local: No external API calls, runs 100% on your machine
• Semi-Closed IE: Uses hardcoded business domain schemas

Prerequisites

• macOS (or Linux/Windows with Ollama support)
• Python 3.8+
• 16GB+ RAM recommended (model is 8.3GB quantized)

Installation

1. Install Ollama

# macOS
brew install ollama
brew services start ollama

# Linux
curl -fsSL https://ollama.com/install.sh | sh

# Windows
# Download from https://ollama.com/download

2. Pull the Model

This downloads the quantized model (8.3GB):

ollama pull jckalo/phi4-ie

3. Install Python Dependency

pip install requests

4. Clone This Repository

git clone https://github.com/JanKalo/phi4-ie-demo.git
cd phi4-ie-demo

Usage

Run the interactive demo:

python interactive_ie_demo.py

Example Session

📝 Enter text:
Apple Inc. acquired Beats Electronics in 2014 for $3 billion.
The company was founded by Dr. Dre and Jimmy Iovine.
###

Type ### on a new line when done, or quit to exit.

Example Output

The model will extract:

Entities:

• Apple Inc. (company)
• Beats Electronics (company)
• 2014 (year)
• $3 billion (financial_metric)
• Dr. Dre (person)
• Jimmy Iovine (person)

Relations:

• acquired
• founded_by

Triples:

• (Apple Inc., acquired, Beats Electronics)
• (Apple Inc., had_revenue_of, $3 billion)
• (Beats Electronics, founded_by, Dr. Dre)
• (Beats Electronics, founded_by, Jimmy Iovine)
• (acquisition, occurred_in_year, 2014)

Performance Notes

⚠️ This is a 15B parameter model running on CPU - inference will be slow:

• First query: 30-60 seconds (model loading into RAM)
• Subsequent queries: 10-30 seconds per extraction
• RAM usage: ~10-12GB during inference
• Best performance: Machines with high-end CPUs or GPUs

Unloading the Model

The model stays loaded in RAM for faster subsequent queries. To free up memory:

# Stop Ollama completely
brew services stop ollama

# Or restart it
brew services restart ollama

Performance Tips

For production use, consider:

• Running on a machine with GPU support
• Using a smaller quantized version (contact maintainer)
• Deploying to a dedicated inference server

Supported Entity Types

company, product, service, industry, brand, country, location, organization, person, founder, position, event, action, founding, acquisition, merger, partnership, expansion, restructuring, divestment, sale, bankruptcy, financial_metric, business_concept, revenue, profit, loss, investment, funding, market_share, competition, market_trend, regulation, innovation, sustainability, corporate_social_responsibility, award, date/time, year, period

Supported Relations

part_of, parent_company_of, subsidiary_of, acquired, divested, owns_brand, holds_stake_in, spun_off, formed_from, founded_by, has_CEO, manufactures, develops, produces, sells, has_product_line, is_a_type_of, is_brand_of, features_technology, uses_material, launched_product, had_revenue_of, had_profit_of, in_year, market_value_of, is_publicly_traded, listed_on_exchange, experienced_growth, occurred_in_year, occurred_on_date, resulted_in, preceded, followed_by, marked_milestone, influenced_by, led_to, was_a_response_to, has_trademark, uses_logo, acquired_trademark_from, is_known_for, launched_campaign

Project Structure

phi4-ie-demo/
├── README.md                    # This file
├── interactive_ie_demo.py       # Main demo script
├── Modelfile                    # (Reference) Ollama model config
└── convert_and_quantize.sh      # (Advanced) Convert HF model to GGUF

Troubleshooting

"Model not found" error

Make sure you've pulled the model:

ollama pull jckalo/phi4-ie

"Cannot connect to Ollama" error

Start the Ollama service:

brew services start ollama

Laptop freezing or very slow

The 15B model is resource-intensive. Try:

• Closing other applications
• Running on a more powerful machine
• Using a smaller model variant

Advanced: Model Conversion

Want to convert a different HuggingFace model to GGUF format? See convert_and_quantize.sh for the conversion pipeline.

This repo includes the conversion script for reference, but you don't need it to run the demo.

Credits

• Base model: Microsoft Phi-4
• Fine-tuned model: FinaPolat/phi4_adaptable_IE
• Quantization: llama.cpp
• Local inference: Ollama
• Model hosting: Ollama Registry

License

Model usage subject to Microsoft Phi-4 license terms.

Citation

If you use this model in your research, please cite:

@misc{phi4-ie-demo,
  author = {Jan Kalo},
  title = {Phi-4 Information Extraction Demo},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/JanKalo/phi4-ie-demo}
}