micro-f1-mask

<div align="center">

ARPA MICRO SERIES: F1 MASK

Zero-Latency PII Scrubbing - 270M Parameter Middleware

<a href="https://huggingface.co/google/gemma-3-270m-it"><img src="https://img.shields.io/badge/base-Gemma_3_270M-bae6fd?style=flat-square" alt="Base Model"></a> <a href="#binary-mapping--tokens"><img src="https://img.shields.io/badge/task-PII_Scrubbing-efcefa?style=flat-square" alt="Task"></a> <a href="#training-methodology"><img src="https://img.shields.io/badge/training-PEFT_LoRA-bbf7d0?style=flat-square" alt="Training"></a> <a href="https://www.apache.org/licenses/LICENSE-2.0.html"><img src="https://img.shields.io/badge/license-Apache_2.0-fde68a?style=flat-square" alt="License"></a>

</div>

---

ARPA Micro Series: F1 Mask is a high-performance fine-tuned model built to provide real-time identification and tokenization of Personally Identifiable Information (PII) for secure cloud computing.

Developed by ARPA Hellenic Logical Systems, it acts as a privacy firewall for incoming/outgoing LLM prompts.

GitHub: arpahls/micro-f1-mask — Full training pipeline, Redis vault, and infrastructure.

Model Summary

F1 Mask is a specialized fine-tune of Gemma 3 270M IT. It is trained exclusively to output structured replace_pii function calls, effectively mapping sensitive data to safe tokens before they reach cloud-based LLMs.

Quick Start

1. Register with Ollama

# Direct SafeTensors registration
ollama create micro-f1-mask --from arpacorp/micro-f1-mask

# Run detection
ollama run micro-f1-mask "John Doe called from 555-0123 about invoice GB29NWBK60161331926819."

2. Python (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("arpacorp/micro-f1-mask")
tokenizer = AutoTokenizer.from_pretrained("arpacorp/micro-f1-mask")

prompt = """<start_of_turn>user
You are Micro F1 Mask. Extract PII and output the 'replace_pii' function call.
Draft an email to Jane Smith at [email protected].<end_of_turn>
<start_of_turn>model
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.0)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Binary Mapping & Tokens

The model uses a deterministic tokenization scheme:

Example Output

{
  "name": "replace_pii",
  "arguments": {
    "entities": [
      {"type": "INDIVIDUAL", "val": "Jane Smith", "id": "[INDIVIDUAL_1]"},
      {"type": "CONTACT", "val": "[email protected]", "id": "[CONTACT_1]"}
    ]
  }
}

Training Methodology

• Dataset: 1,000 synthetic samples generated via high-entropy LLM workflows.
• Method: PEFT / LoRA (Rank 16, Alpha 32).
• Epochs: 3.
• Accuracy: 76.10% (token-level generation).
• Latency: Sub-50ms (inference on RTX 2070).

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 an evaluation script to benchmark 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 structure, generate 500 synthetic variations of that specific edge-case, and re-run the fine-tuning pipeline.

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

For mission-critical data, we recommend extending the middleware 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 application UI highlights the detected entities to the user. The user manually verifies the masking before the payload is authorized to hit the external cloud.
• Post-Reconstruction Review (Quality Control): Allow the fully automated process to finish. Before the final reconstructed cloud response is saved or emailed, route it to an analyst dashboard where a human can manually verify the grammar of the reconstructed payload.

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]

Ethical Considerations

Data Provenance: No real PII was used in the training of this model. All examples were synthetically generated to mimic enterprise communication patterns.

Intended Use: This model is designed for middleware. It is not intended to be used as a conversational assistant. It is a one-way security gate that focuses exclusively on privacy enforcement.

---

<div align="center"> <img src="https://raw.githubusercontent.com/ARPAHLS/skillware/main/assets/arpalogo.png" width="50px">

Built by ARPA Hellenic Logical Systems

</div>