Hugging Face's logo

Minibase
/
DeId-Small

GGUF
English
text-deidentification
privacy
pii-removal
text2text-generation
medical
legal
hr
llama
minibase
small-model
2048-context
Eval Results
Model card Files
xet
Community
DeId-Small / README.md
Minibase's picture
Minibase
Upload README.md with huggingface_hub
e322de3 verified
preview code
|
raw
history blame
11.5 kB
metadata
language:
  - en
tags:
  - text-deidentification
  - privacy
  - pii-removal
  - text2text-generation
  - medical
  - legal
  - hr
  - llama
  - gguf
  - minibase
  - small-model
  - 2048-context
license: apache-2.0
datasets:
  - custom
metrics:
  - completeness-score
  - pii-detection-rate
  - semantic-preservation
  - latency
model-index:
  - name: DeId-Small
    results:
      - task:
          type: text-deidentification
          name: Completeness Score
        dataset:
          type: custom
          name: Personal De-identifier Benchmark
          config: mixed-domains
          split: test
        metrics:
          - type: completeness-score
            value: 0.64
            name: Overall Completeness
          - type: pii-detection-rate
            value: 0.203
            name: PII Detection Rate
          - type: semantic-preservation
            value: 0.109
            name: Semantic Preservation
          - type: latency
            value: 492.4
            name: Average Latency (ms)

DeId-Small 🤖

A compact, privacy-focused text de-identification model for removing personal identifiers while preserving meaning.

Model Size Architecture Context Window License Discord

Built by Minibase - Train and deploy small AI models from your browser. Browse all of the models and datasets available on the Minibase Marketplace.

📋 Model Summary

Minibase-DeId-Small is a specialized language model fine-tuned for text de-identification tasks. It automatically detects and replaces personal identifiers (PII) such as names, dates, addresses, phone numbers, and other sensitive information with standardized placeholder tags while preserving the original meaning and context of the text.

Key Features

  • 🔒 Privacy-First: Removes personal identifiers automatically
  • 🎯 High Completeness: 64% of texts fully de-identified
  • 📏 Compact Size: 136MB (Q8_0 quantized)
  • ⚡ Fast Inference: ~492ms average response time
  • 🌐 Multi-Domain: Works across medical, legal, HR, and general text
  • 🔄 Local Processing: No data sent to external servers

🚀 Quick Start

Local Inference (Recommended)

  1. Install llama.cpp (if not already installed):

    git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp && make

  2. Download and run the model:

    # Download model files
wget https://huggingface.co/Minibase/DeId-Small/resolve/main/model.gguf
wget https://huggingface.co/Minibase/DeId-Small/resolve/main/deid_inference.py

# Make executable and run
chmod +x run_server.sh
./run_server.sh

  3. Make API calls:

    import requests

# De-identify text
response = requests.post("http://127.0.0.1:8000/completion", json={
    "prompt": "Instruction: De-identify this text by replacing all personal information with placeholders.\n\nInput: Patient John Smith, born 1985-03-15, lives at 123 Main St.\n\nResponse: ",
    "max_tokens": 256,
    "temperature": 0.1
})

result = response.json()
print(result["content"])  # "Patient [FIRSTNAME_1] [LASTNAME_1], born [DOB_1], lives at [BUILDINGNUMBER_1] [STREET_1]."

Python Client 

from deid_inference import DeIdClient

# Initialize client
client = DeIdClient()

# De-identify text
sensitive_text = "Dr. Sarah Johnson called from (555) 123-4567 about patient Michael Brown."
clean_text = client.deidentify_text(sensitive_text)

print(clean_text)  # "Dr. [FIRSTNAME_1] [LASTNAME_1] called from [PHONE_1] about patient [FIRSTNAME_2] [LASTNAME_2]."

📊 Benchmarks & Performance

Overall Performance (100 samples)

Metric Score Description
Completeness Score 64.0% Percentage of texts fully de-identified
PII Detection Rate 20.3% How well expected PII patterns are detected
Semantic Preservation 10.9% How well original meaning is preserved
Average Latency 492ms Response time performance

Domain-Specific Performance

Domain Samples Completeness Best For
General Text 40 75.0% Mixed content, everyday text
Customer Service 6 100% Support interactions, inquiries
Medical Records 33 60.6% Patient notes, clinical data
Legal Documents 6 50.0% Depositions, contracts
HR Records 11 27.3% Employee data, personnel files
Research Data 4 50.0% Study participants, academic data

Performance Insights:

  • ✅ Excellent at customer service (100% completeness)
  • ✅ Strong general-purpose performance (75% completeness)
  • ✅ Good medical record handling (60.6% completeness)
  • ⚠️ Challenging with structured HR data (27.3% completeness)

🏗️ Technical Details

Model Architecture

  • Architecture: LlamaForCausalLM
  • Parameters: 135M (small capacity)
  • Context Window: 2,048 tokens
  • Max Position Embeddings: 2,048
  • Quantization: GGUF (Q8_0 quantization)
  • File Size: 136MB
  • Memory Requirements: 8GB RAM minimum, 16GB recommended

Training Details

  • Base Model: Custom-trained Llama architecture
  • Fine-tuning Dataset: Curated PII-parallel text pairs
  • Training Objective: Instruction-following for de-identification
  • Optimization: Quantized for efficient inference
  • Model Scale: Small capacity optimized for speed

System Requirements

Component Minimum Recommended
Operating System Linux, macOS, Windows Linux or macOS
RAM 8GB 16GB
Storage 150MB free space 500MB free space
Python 3.8+ 3.10+
Dependencies llama.cpp llama.cpp, requests

Notes:

  • ✅ CPU-only inference supported but slower
  • ✅ GPU acceleration provides significant speed improvements
  • ✅ Apple Silicon users get Metal acceleration automatically

📖 Usage Examples

Basic De-identification 

# Input: "John Smith from New York called about his account."
# Output: "[FIRSTNAME_1] [LASTNAME_1] from [CITY_1] called about his account."

# Input: "Patient born on 1990-05-15 visited Dr. Williams."
# Output: "Patient born on [DOB_1] visited Dr. [LASTNAME_1]."

Medical Records 

# Input: "Sarah Johnson, DOB 05/12/1980, visited St. Jude Hospital."
# Output: "[FIRSTNAME_1] [LASTNAME_1], DOB [DOB_1], visited [HOSPITAL_1]."

# Input: "Dr. Michael Brown called from (555) 123-4567."
# Output: "Dr. [FIRSTNAME_1] [LASTNAME_1] called from [PHONE_1]."

Legal Documents 

# Input: "Attorney Robert Davis from Legal Eagles LLP filed the motion."
# Output: "Attorney [FIRSTNAME_1] [LASTNAME_1] from [ORGANIZATION_1] filed the motion."

# Input: "Case LD-2022-007 was filed on December 1, 2022."
# Output: "Case [CASE_ID_1] was filed on [DATE_1]."

HR Records 

# Input: "Employee John Doe earns $85,000 annually."
# Output: "Employee [FIRSTNAME_1] Doe earns [CURRENCYSYMBOL_1][AMOUNT_1] annually."

# Input: "Contact [email protected] for details."
# Output: "Contact [EMAIL_1] for details."

🔧 Advanced Configuration

Server Configuration 

# GPU acceleration (macOS with Metal)
llama-server \
  -m model.gguf \
  --host 127.0.0.1 \
  --port 8000 \
  --n-gpu-layers 35 \
  --ctx-size 2048 \
  --metal

# CPU-only (higher memory usage)
llama-server \
  -m model.gguf \
  --host 127.0.0.1 \
  --port 8000 \
  --n-gpu-layers 0 \
  --threads 8 \
  --ctx-size 2048

Temperature Settings

Temperature Range Approach Description
0.0-0.2 Conservative (Recommended) Precise, consistent de-identification
0.3-0.5 Balanced Good balance of accuracy and flexibility
0.6-1.0 Creative More flexible but may miss some PII

📚 Limitations & Biases

Current Limitations

Limitation Description Impact
Placeholder Format Uses specific naming conventions (e.g., [FIRSTNAME_1]) May not match all expected formats
Complex Contexts May struggle with highly nested or ambiguous PII Could miss subtle personal information
Language Scope Primarily trained on English text Limited performance on other languages
Context Window Limited to 2,048 token context window Cannot process very long documents
Structured Data Less effective on highly formatted data (tables, forms) Lower performance on structured HR/financial data

Potential Biases

Bias Type Description Mitigation
Cultural Names May not recognize all international naming patterns Regular updates with diverse data
Regional Formats Limited exposure to regional address/phone formats Expand training data coverage
Emerging PII May not recognize newest types of personal data Continuous model updates
Domain Specificity Performance varies across different text types Use domain-specific fine-tuning

🤝 Contributing

We welcome contributions! Please see our Contributing Guide for details.

Development Setup 

# Clone the repository
git clone https://github.com/minibase-ai/deid-small
cd deid-small

# Install dependencies
pip install -r requirements.txt

# Run tests
python -m pytest tests/

📜 Citation

If you use DeId-Small in your research, please cite:

@misc{deid-small-2025,
  title={DeId-Small: A Compact Text De-identification Model},
  author={Minibase AI Team},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/Minibase/DeId-Small}
}

📞 Contact & Community

Support

📋 License

This model is released under the Apache License 2.0.

🙏 Acknowledgments

  • Personal De-identifier Benchmark Dataset: Used for training and evaluation
  • llama.cpp: For efficient local inference
  • Hugging Face: For model hosting and community
  • Our amazing community: For feedback and contributions

Built with ❤️ by the Minibase team

Making AI safer and more accessible for everyone

🌟 Star us on GitHub • 📖 Read the docs • 💬 Join our Discord