micro-f1-mask

# ARPA MICRO SERIES: F1 MASK **Zero-Latency PII Scrubbing Middleware for Enterprise Cloud Compliance** Mission • Documentation • Quick Start • Data Generation • PII Categories • Deployment • Contact

Mission

ARPA Micro Series: F1 Mask is a purpose-built privacy bridge designed to intercept outgoing LLM prompts, detect and tokenize Personally Identifiable Information (PII) using a high-efficiency 270M parameter model, and forward only sanitized content to cloud APIs.

The core philosophy is Privacy by Architecture: sensitive data never leaves your infrastructure.

"Privacy is not something that I'm merely entitled to, it's an absolute prerequisite." — Marlon Brando

Model Weights: arpacorp/micro-f1-mask
Source Code: Training pipeline, middleware proxy, Redis vault, and specialized deployment tools.

Documentation

Full technical reference library:

• Training Specification: Hardware, hyperparameters, and convergence logs.
• Testing & QA Guide: Unit and scenario test execution.
• API Reference: FastAPI Bridge endpoint and payload schemas.
• Evaluation & Ethics: Accuracy, latency, and data provenance.

Architecture

┌──────────────────────────────────────────────────────────────────┐
│                        CLIENT / AGENT                            │
│                     "Call John Doe at 555-0123"                   │
└────────────────────────────┬─────────────────────────────────────┘
                             │
                             ▼
┌──────────────────────────────────────────────────────────────────┐
│  PHASE A: THE MASK                                               │
│  ┌─────────────┐    ┌───────────────┐    ┌────────────────────┐  │
│  │  F1 Mask    │───▶│  PII Detect   │───▶│  Redis Vault       │  │
│  │  (Ollama)   │    │  replace_pii  │    │  John → [IND_1]    │  │
│  └─────────────┘    └───────────────┘    │  555  → [CON_1]    │  │
│                                          └────────────────────┘  │
│  Scrubbed: "Call [INDIVIDUAL_1] at [CONTACT_1]"                  │
└────────────────────────────┬─────────────────────────────────────┘
                             │
                             ▼
┌──────────────────────────────────────────────────────────────────┐
│  PHASE B: THE CLOUD                                              │
│  GPT-4 / Claude / Gemini receives ONLY tokenized text.           │
│  No real PII leaves your network.                                │
└────────────────────────────┬─────────────────────────────────────┘
                             │
                             ▼
┌──────────────────────────────────────────────────────────────────┐
│  PHASE C: THE REVEAL                                             │
│  Cloud response tokens are reconstructed from the Redis Vault.   │
│  "[INDIVIDUAL_1] prefers email" → "John Doe prefers email"       │
└──────────────────────────────────────────────────────────────────┘

Quick Start

1. Installation

git clone https://github.com/arpahls/micro-f1-mask.git
cd micro-f1-mask

# Install PyTorch with CUDA support
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

# Install project dependencies
pip install -r requirements.txt

# Configure environment
cp .env.example .env

2. Model Registration (Ollama)

# Register the model from the local merged directory
ollama create micro-f1-mask -f Ollama.Modelfile

# Optional: Register with 4-bit quantization for performance
ollama create micro-f1-mask --quantize q4_K_M -f Ollama.Modelfile

3. Start Infrastructure

# Start Redis Vault
docker compose up -d

# Start F1 Mask Bridge (FastAPI Proxy)
python micro_f1_mask_bridge.py

Data Generation

To maintain the privacy of the original 1000-sample dataset, we provide a high-entropy synthetic generator. Users can create their own training data using the provided script.

Configuration

Modify synthetic_generator.py to adjust:

• Model: Default is gemini-2.0-flash-lite (high speed/low cost).
• Quantity: Change the loop range to generate thousands of unique scenarios.
• Diversity: Adjust the system prompt to introduce specific industry jargon or PII formats.

# Generate a new synthetic dataset
python synthetic_generator.py

Production Optimization Roadmap

While this repository provides a fully functional 1,000-sample prototype, reaching 95%+ enterprise accuracy requires the following architectural optimizations:

1. Hard-Negative Mining (Re-training)

To push accuracy into the high 90s, the model must iteratively learn from its mistakes:

1. Scale: Use the synthetic generator to produce 10,000 - 50,000 highly diverse samples tailored to your industry.
2. Evaluate: Run evaluation.py continuously against samples of your real-world traffic.
3. Mine Edge Cases: Every time the model misses a PII token (a "hard negative"), extract that sentence, generate 500 synthetic variations of that specific edge-case structure, and re-run the LoRA fine-tuning pipeline.

2. Human-In-The-Loop (HITL) Workflows

For mission-critical data, we recommend extending the FastAPI Bridge to include human oversight:

• Pre-Cloud Quarantine (Maximum Security): Modify the endpoint so that when F1 Mask detects PII, the API payload pauses. The UI highlights the detected entities to the user (e.g., "We found [John Doe]. Click to approve."). The user manually verifies the masking before it is authorized to hit the cloud.
• Post-Reconstruction Review (Quality Control): Allow the fully automated process to finish. Before the final reconstructed cloud response is saved to a database or emailed to a client, route it to an admin dashboard where a human can manually verify the reconstructed grammar.

PII Categories

F1 Mask detects and tokenizes 6 core entity categories:

Deployment

Multi-Turn Session Management

The Redis Vault ensures that PII mappings remain consistent across a conversation:

• Idempotency: The same PII value always yields the same token within a session.
• Isolation: Each session_id has a unique, protected namespace.
• Lifecycle: Sessions auto-expire after a configurable TTL (default: 2 hours).

Production Checklist

• [ ] Enable Redis password authentication.
• [ ] Configure TLS for the Bridge API endpoint.
• [ ] Set log level to WARNING to prevent PII leakage in system logs.
• [ ] Ensure the model is running on local hardware for maximum security.

Ethics & Responsible Use

Design Principles

1. Privacy by Architecture: Data protection is enforced at the network layer.
2. Minimalist Scope: 270M parameters focused strictly on scrubbing, avoiding general-purpose LLM risks.
3. Synthetic Provenance: The model was trained exclusively on synthetic data. No real human data was used.

Usage Disclosure

F1 Mask is designed for enterprise data protection and compliance. It should not be used as a substitute for comprehensive security audits or as a primary defense against targeted adversarial extraction.

Enterprise Solutions

The public release of ARPA F1 Mask serves as a lightweight demonstration of how the Function One (F1) architecture can be fine-tuned for structured privacy enforcement.

For mission-critical infrastructure, ARPA offers an actively maintained, highly robust enterprise tier. Organizations can deploy our gated version out-of-the-box and completely offload the burden of continuous maintenance, bespoke fine-tuning, concept drift avoidance, and rigorous scenario evaluations.

For enterprise licensing or to discuss tailoring the F1 model to your proprietary data schemas, reach out to: [email protected]

Contact

• Commercial: [email protected]
• Security: [email protected]
• Issues: GitHub Issues
• Organization: ARPA Hellenic Logical Systems

Built & Maintained by ARPA Hellenic Logical Systems