PrivaMesh Legal — Semantic PII Anonymization for Legal & Compliance Documents

Table of Contents

• Overview
• Key Differentiators vs. Existing Approaches
• The PrivaMesh Framework
• Supported Privacy Categories
• Quick Start
• Advanced Usage
• Model Architecture
• Training Details
• Evaluation & Benchmarks
• Deployment
• Regulatory Coverage
• Limitations & Risks
• Citation
• License

Overview

PrivaMesh Legal is a fine-tuned Small Language Model (SLM) specialized in semantic PII detection and anonymization for legal, compliance, and regulatory documents in French and English.

It is designed for:

• Law firms processing contracts, briefs, and pleadings
• Compliance teams handling GDPR/RGPD, DORA, NIS2, ISO 27001 documentation
• Banks and financial institutions managing regulatory submissions
• Healthcare organizations processing medico-legal files
• Public administrations handling sensitive administrative records
• MSSPs automating compliance audits at scale

What makes PrivaMesh Legal different

Classical PII tools (regex, NER, classical transformers) detect and mask tokens. They answer: "Is this token a person's name?"

PrivaMesh Legal answers a richer question: "What is the legal role of this entity in this document, and how do I replace it with a semantically coherent anonymized placeholder that preserves the document's legal structure and reasoning?"

Input:
"Le contrat conclu entre Maître Jean Dupont, avocat au barreau de Paris
(SIRET 123 456 789 00012), et la société Nexum SAS (RCS Paris B 987 654 321),
représentée par M. Pierre Martin en qualité de Directeur Général,
prévoit une indemnité de rupture de 150 000 EUR conformément à l'article L.1237-19 du Code du travail."

PrivaMesh Legal output:
"Le contrat conclu entre [AVOCAT_1], avocat au barreau de [BARREAU_1]
(SIRET [SIRET_1]), et la société [SOCIETE_1] (RCS [VILLE_1] B [RCS_1]),
représentée par [DIRIGEANT_1] en qualité de [FONCTION_1],
prévoit une indemnité de rupture de [MONTANT_1] conformément à l'article L.1237-19 du Code du travail."

Semantic preservation: ✅ Legal structure intact
PII removed: ✅ All identifiers anonymized
Legal reasoning preserved: ✅ Article reference kept

Key Differentiators vs. Existing Approaches

The PrivaMesh Framework

PrivaMesh Legal is one node in the PrivaMesh collaborative multi-SLM mesh. Each node is a specialized SLM fine-tuned on a specific domain. An orchestrator agent coordinates them at inference time.

Upcoming PrivaMesh models:

Supported Privacy Categories

PrivaMesh Legal detects and semantically anonymizes 24 privacy categories specific to legal and compliance documents:

Natural Persons

Legal Entities

Financial & Contractual

Contact & Location

Temporal & Reference

Regulatory & Compliance Specific

Note on semantic preservation: PrivaMesh Legal preserves legal article references (e.g., article L.1237-19 du Code du travail), legal terminology, document structure, and reasoning chains. Only identifiers and personal data are anonymized.

Quick Start

Installation

pip install transformers torch privamesh

Basic usage — Pipeline API

from privamesh import PrivaMeshLegal

# Initialize (runs fully on-premise, no API call)
model = PrivaMeshLegal.from_pretrained("privamesh/privamesh-legal")

# Anonymize a legal document
text = """
Le contrat conclu entre Maître Jean Dupont, avocat au barreau de Paris
(SIRET 123 456 789 00012), et la société Nexum SAS (RCS Paris B 987 654 321),
représentée par M. Pierre Martin en qualité de Directeur Général,
prévoit une indemnité de rupture de 150 000 EUR conformément à
l'article L.1237-19 du Code du travail.
"""

result = model.anonymize(text)

print(result.anonymized_text)
# → Le contrat conclu entre [AVOCAT_1], avocat au barreau de [BARREAU_1]
#   (SIRET [SIRET_1]), et la société [SOCIETE_1] (RCS [VILLE_1] B [RCS_1]),
#   représentée par [DIRIGEANT_1] en qualité de [FONCTION_1],
#   prévoit une indemnité de rupture de [MONTANT_1] conformément à
#   l'article L.1237-19 du Code du travail.

print(result.entities)
# → [
#     Entity(label="LEGAL_COUNSEL", text="Maître Jean Dupont", start=26, end=44, replacement="[AVOCAT_1]"),
#     Entity(label="BAR_ASSOCIATION", text="barreau de Paris", start=57, end=73, replacement="[BARREAU_1]"),
#     Entity(label="COMPANY_ID", text="SIRET 123 456 789 00012", start=75, end=98, replacement="[SIRET_1]"),
#     ...
#   ]

print(result.semantic_score)
# → 0.94  (BERTScore semantic preservation)

print(result.privacy_recall)
# → 0.97  (fraction of PII entities detected)

Using with HuggingFace Transformers directly

from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("privamesh/privamesh-legal")
model = AutoModelForTokenClassification.from_pretrained(
    "privamesh/privamesh-legal",
    device_map="auto"
)

text = "Le contrat signé par Jean Dupont le 15 mars 2024."
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model(**inputs)

predicted_ids = outputs.logits.argmax(dim=-1)
predicted_labels = [
    model.config.id2label[id.item()]
    for id in predicted_ids[0]
]
print(predicted_labels)

Advanced Usage

Batch processing — high throughput

from privamesh import PrivaMeshLegal

model = PrivaMeshLegal.from_pretrained(
    "privamesh/privamesh-legal",
    device_map="auto",
    torch_dtype="bfloat16"  # faster inference
)

documents = [doc1, doc2, doc3, ...]  # list of strings

results = model.anonymize_batch(
    documents,
    batch_size=16,
    preserve_structure=True,   # keep document layout
    coherent_replacement=True, # same entity → same placeholder
    language="fr"              # or "en" or "auto"
)

Precision / Recall tuning

result = model.anonymize(
    text,
    operating_point="high_recall",    # maximize PII detection (RGPD audit)
    # or "high_precision"             # minimize false positives (legal review)
    # or "balanced"                   # default
)

Custom label policy — fine-grained control

# Anonymize only specific categories
result = model.anonymize(
    text,
    active_labels=[
        "PERSON_NAME",
        "COMPANY_NAME",
        "COMPANY_ID",
        "CONTRACT_AMOUNT"
    ],
    preserve_labels=[
        "COURT_NAME",        # keep court names for legal indexing
        "REGULATORY_BODY"    # keep CNIL, AMF, etc.
    ]
)

Consistent anonymization across a document set

# Anonymize a full case file — same entity gets same placeholder across all docs
from privamesh import PrivaMeshLegal, AnonymizationContext

ctx = AnonymizationContext()  # shared entity registry

contract = model.anonymize(contract_text, context=ctx)
brief     = model.anonymize(brief_text,   context=ctx)
judgment  = model.anonymize(judgment_text, context=ctx)

# "Jean Dupont" → "[PERSONNE_1]" consistently across all three documents

PrivaMesh multi-SLM orchestration

from privamesh import PrivaMeshOrchestrator

# Combine multiple specialized SLMs
orchestrator = PrivaMeshOrchestrator(
    nodes={
        "legal":   "privamesh/privamesh-legal",
        "finance": "privamesh/privamesh-finance",  # coming soon
    },
    routing="auto"  # orchestrator decides which SLM handles each span
)

# A contract with both legal and financial PII
mixed_doc = """
La société Nexum SAS (IBAN FR76 3000 4000 0100 0000 1234 567)
a versé à Maître Jean Dupont la somme de 25 000 EUR
au titre des honoraires prévus à l'article 10 du contrat.
"""

result = orchestrator.anonymize(mixed_doc)

Model Architecture

PrivaMesh Legal is built on a fine-tuned Mistral-Small-3.1 backbone — a French-native, Apache 2.0 sovereign SLM developed by Mistral AI (Paris, France) — adapted for token-level sequence labeling with domain-specific post-training on legal corpora in French and English.

Why Mistral? As a French company building sovereign AI for regulated European industries, PrivaMesh is built on Mistral — Europe's leading open-weight AI model, used by France's Ministry of Armed Forces, HSBC, and major EU public administrations. This is not just a technical choice — it is a sovereignty statement.

Architecture overview

Base model    : mistralai/Mistral-Small-3.1 (Apache 2.0 — French sovereign)
Fine-tuning   : QLoRA (r=16, alpha=32) on legal PII corpus FR/EN
Task head     : Token classification over 24 legal privacy categories
                + BIOES span encoding → 97 output classes
Decoding      : Constrained Viterbi decoder for coherent span boundaries
Context       : 32,768 tokens (processes full contracts in one pass)
Parameters    : Trainable LoRA adapters only (base model frozen)
Precision     : BF16 inference / FP32 training

Label encoding — BIOES scheme

Each of the 24 privacy categories is encoded in BIOES format:

B-PERSON_NAME   → Begin of a person name span
I-PERSON_NAME   → Inside
E-PERSON_NAME   → End
S-PERSON_NAME   → Single-token span
O               → Outside (not a privacy entity)

Total output classes: 1 (O) + 24 categories × 4 (BIOES) = 97 classes

Semantic replacement strategy

Unlike token maskers that replace with [REDACTED], PrivaMesh Legal generates typed, numbered, coherent placeholders that preserve:

1. Entity type — [AVOCAT_1] vs [SOCIETE_1] vs [MONTANT_1]
2. Entity role — the legal function is encoded in the placeholder type
3. Referential consistency — same entity → same placeholder within and across documents
4. Grammatical agreement — French gendered replacements (coming in v1.1)

Training Details

Base model

Training data

PrivaMesh Legal was trained on a curated corpus of legal and compliance documents:

Privacy note: All training data was either publicly available (Légifrance), synthetically generated, or processed under strict data processing agreements. No real personal data was retained in model weights.

Data augmentation

To improve robustness, training data was augmented with:

• Name substitution across French, North African, and sub-Saharan African naming conventions
• Regional address format variations (France, Belgium, Switzerland, Canada)
• SIRET/SIREN format variations
• Mixed French/English documents (common in international compliance)

Evaluation & Benchmarks

Key metrics at a glance

Benchmark 1 — PII detection F1 across tools

Benchmark 2 — Semantic preservation (BERTScore)

Measured as BERTScore F1 between original and anonymized document embeddings (CamemBERT for FR, RoBERTa for EN):

Benchmark 3 — F1 per PII category

Benchmark 4 — Training loss curve (QLoRA fine-tuning)

Benchmark 5 — Precision / Recall tradeoff

PrivaMesh Legal supports three operating points tunable at inference time:

Benchmark 6 — Throughput vs document length

Benchmarked on a single A10G GPU (24GB):

Deployment

On-premise deployment (recommended)

PrivaMesh Legal is designed for sovereign, on-premise deployment. No data leaves your infrastructure.

# Pull model locally
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="privamesh/privamesh-legal",
    local_dir="./models/privamesh-legal",
    ignore_patterns=["*.msgpack", "*.h5"]
)

# Load from local path — fully air-gapped
from privamesh import PrivaMeshLegal

model = PrivaMeshLegal.from_pretrained(
    "./models/privamesh-legal",
    device_map="auto",
    local_files_only=True  # no internet connection required
)

Hardware requirements

Quantized versions

# 4-bit quantization — runs on 8GB VRAM
from transformers import BitsAndBytesConfig

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16
)

model = PrivaMeshLegal.from_pretrained(
    "privamesh/privamesh-legal",
    quantization_config=bnb_config,
    device_map="auto"
)

Docker deployment

FROM python:3.11-slim

RUN pip install privamesh transformers torch

COPY ./models/privamesh-legal /models/privamesh-legal

EXPOSE 8080

CMD ["privamesh", "serve", "--model", "/models/privamesh-legal", "--port", "8080"]

docker build -t privamesh-legal .
docker run -p 8080:8080 --gpus all privamesh-legal

REST API (built-in server)

privamesh serve --model privamesh/privamesh-legal --port 8080

curl -X POST http://localhost:8080/anonymize \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Le contrat signé par Jean Dupont le 15 mars 2024.",
    "language": "fr",
    "operating_point": "high_recall"
  }'

Regulatory Coverage

PrivaMesh Legal is designed to support compliance with the following regulatory frameworks:

Limitations & Risks

Known limitations

1. Language coverage PrivaMesh Legal is optimized for French and English. Performance may degrade on other languages, mixed-language documents with code-switching, or heavily technical jargon outside the training distribution.

2. Rare naming conventions Detection performance may be lower for names following naming conventions underrepresented in training data (some regional French dialects, transliterated names, highly abbreviated forms).

3. Implicit PII PrivaMesh Legal detects explicit PII. Implicit or inferred PII (e.g., identifying someone from their unique job description without naming them) is not in scope and requires additional processing layers.

4. Dynamic label policies Like openai/privacy-filter, changing which categories are anonymized requires fine-tuning rather than runtime configuration (except for the active_labels filter, which suppresses labels post-detection).

5. Not a legal guarantee PrivaMesh Legal is a technical anonymization aid. It does not constitute legal advice or a guarantee of RGPD compliance. Human review is recommended for high-stakes workflows.

Risk: Over-reliance

Do not use PrivaMesh Legal as your sole anonymization layer for high-sensitivity documents. It is designed as a primary processing layer in a privacy-by-design architecture that includes human review, audit trails, and access controls.

Responsible use

PrivaMesh Legal is intended for data protection and privacy-preserving AI workflows. It must not be used to:

• Circumvent legitimate legal discovery or regulatory oversight
• Process data without appropriate legal basis
• Bypass consent mechanisms required under RGPD

Citation

If you use PrivaMesh Legal in your research or production systems, please cite:

@misc{privamesh2026legal,
  title     = {PrivaMesh: A Collaborative Multi-SLM Framework for Semantic Data Anonymization in Sovereign Agentic AI Pipelines},
  author    = {Sabri ALLANI et Ahmed HERSI},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/sallani/PrivaMesh},
  note      = {PrivaMesh Legal — Domain-specialized SLM for legal and compliance document anonymization. Base model: Mistral-Small-3.1 (Apache 2.0)}
}

📄 Paper: "PrivaMesh: A Collaborative Multi-SLM Framework for Semantic Data Anonymization in Sovereign On-Premise Agentic AI Pipelines" — preprint submission arXiv 2026, Q1 journal under review.

Contributing

PrivaMesh is an open research initiative. Contributions welcome:

• 🐛 Report issues
• 📊 Share evaluation results
• 🔧 Contribute to the framework
• 📝 Request new domains

License

Apache 2.0 — Free for research, experimentation, and commercial deployment.

Built on Mistral-Small-3.1 (Apache 2.0) by Mistral AI, Paris 🇫🇷

See LICENSE for full terms.