The AI Layer That Makes Sense Unbeatable
18 specialized models. 5 detection layers. Every model runs exclusively on Segmento's private infrastructure — your data never reaches a third-party AI provider.
By The Numbers
An ensemble built for precision
Model Stack
18 models. 5 detection layers.
Each sub-layer targets a specific challenge. Together, they form an ensemble no single model can match.
4A — Primary NER: ModernBERT & DeBERTa models for high-accuracy Named Entity Recognition
joneauxedgar/pasteproof-pii-detector-v2
ModernBERT-base (149M)
ModernBERT-base with 149M parameters. Covers 27 PII types across PCI/HIPAA/GDPR frameworks. Trained on 150K synthetic examples with BIO tagging.
llm-semantic-router/mmbert32k-pii-detector-merged
ModernBERT 307M + YaRN (32K context)
307M parameter ModernBERT with YaRN extending context to 32K tokens. Unmatched open-source coverage for long-form legal and compliance documents.
OpenMed PII Family (FR/DE/IT variants)
ModernBERT-large (395M)
ModernBERT-large (395M) covering 55+ EU-localized PII types including French NSS and Italian Codice Fiscale. Multilingual healthcare focus.
iiiorg/piiranha-v1-detect-personal-information
DeBERTa-v3-base
DeBERTa-v3-base with 99.44% binary accuracy. The fastest deployed model at 25ms GPU. Ideal as a high-speed first-pass gate across 6 languages.
exdsgift/NerGuard-0.3B
mDeBERTa-v3-base (0.3B)
mDeBERTa-v3-base with the highest throughput in the 4A sub-layer. 33ms median latency across 8 EU languages. Ideal for real-time pipelines.
lakshyakh93/deberta_finetuned_pii
DeBERTa-v3-base
DeBERTa-v3-base fine-tuned on ai4privacy/pii-masking-300k. Serves as a warm fallback and general-purpose baseline in the 4A ensemble.
Technical Comparison
Every model. Every metric.
Filter and sort across the full stack. Click column headers to re-rank by accuracy or speed.
Showing 18 models
| Model | Layer | Architecture | Top Metric | Context Window | Best For | Trainable | Latency |
|---|---|---|---|---|---|---|---|
joneauxedgar/pasteproof-pii-detector-v2 In Use | 4A | ModernBERT-base (149M) | F1 0.970 F1 (held-out) | 8,192 tokens | Long compliance docs, leakage prevention, intentional variation coverage | Yes | ~120ms GPU |
llm-semantic-router/mmbert32k-pii-detector-merged In Use | 4A | ModernBERT 307M + YaRN (32K context) | F1 0.969 F1 (reported) | 32,768 tokens | Extreme-length documents, legal contracts, batch reports with dense PII | Yes | ~400ms GPU |
OpenMed PII Family (FR/DE/IT variants) | 4A | ModernBERT-large (395M) | >F1 0.960 F1 (reported) | 8,192 tokens | EU multilingual (FR/DE/IT), GDPR localized entity formats, healthcare records | Yes | ~180ms GPU |
iiiorg/piiranha-v1-detect-personal-information In Use | 4A | DeBERTa-v3-base | 99.44% acc · F1 0.931 Binary Acc / F1-macro | 512 tokens (sub-256 optimal) | High-speed short-segment screening, multilingual real-time API, binary PII gate | Yes | ~25ms GPU |
exdsgift/NerGuard-0.3B | 4A | mDeBERTa-v3-base (0.3B) | F1-macro 0.9963 F1-macro (in-dist) | 512 tokens | Ultra-low latency EU multilingual, high-throughput real-time pipelines | Yes | 33ms GPU |
lakshyakh93/deberta_finetuned_pii In Use | 4A | DeBERTa-v3-base | F1 ~0.920 F1 (est.) | 512 tokens | General-purpose PII baseline, interpretable benchmark | Yes | ~30ms GPU |
knowledgator/gliner-pii-large-v1.0 | 4B | GLiNER-large (bi-encoder) | F1 0.833 · Prec 0.874 F1 / Precision | 512 tokens | Minimizing false positives, broadest entity coverage, production compliance audits | Yes | ~45ms GPU |
nvidia/gliner-PII-0.1 In Use | 4B | GLiNER DeBERTa (570M) | Strict F1 0.870 Strict F1 | 512 tokens | Enterprise compliance, healthcare / finance / legal tri-domain | Yes | ~60ms GPU |
gretelai/gretel-gliner-bi-large-v1.0 | 4B | GLiNER-large (bidirectional) | F1 0.950 F1 (internal bench) | 512 tokens | Dual PII+PHI detection in one pass, HIPAA + GDPR simultaneously | Yes | ~50ms GPU |
OvermindLab/nerpa | 4B | GLiNER2 (unified NER + structured extraction) | Micro-Prec 0.930 Micro-Precision | 512 tokens | Disambiguation of overlapping entity types, beats AWS Comprehend | Yes | ~55ms GPU |
urchade/gliner_small-v2.1 In Use | 4B | GLiNER-small (DeBERTa-v3-small encoder) | F1 ~0.850 F1 (general NER) | 512 tokens | Zero-shot custom entities, prototyping new PII types, ultra-fast inference | Yes | ~15ms GPU |
Surya OCR (Datalab) | 4C | Detection + segmentation models | Prec 0.99 · Rec 0.96 Table Det. Prec / Recall | Full page canvas | Scanned document pre-processing, reading order correction, bounding box extraction | Partial | ~620ms/page GPU |
nielsr/layoutlmv3-finetuned-cord | 4C | LayoutLMv3-base (multimodal) | F1 0.9638 F1 (CORD) | 512 tokens + image patches | Receipt & invoice spatial PII extraction, structured financial document parsing | Yes | ~200ms GPU (incl. OCR) |
nielsr/layoutlmv3-finetuned-funsd | 4C | LayoutLMv3-base (multimodal) | F1 0.9078 F1 (FUNSD) | 512 tokens + image patches | Scanned form key-value extraction, insurance/government forms | Yes | ~200ms GPU (incl. OCR) |
parthesh111/layoutlmv3-finetune-bioes-new | 4C | LayoutLMv3-base + PaddleOCR | F1 ~0.920 F1 (medical lab reports) | 512 tokens + image patches | Scanned medical lab report de-identification, HIPAA PHI spatial extraction | Yes | ~250ms GPU (incl. OCR) |
fast-langdetect (FastText lite) | 4D | FastText (bag-of-n-grams classifier) | ~98% (common langs) Top-1 Accuracy (common langs) | Sentence/paragraph | Edge-level language routing, <1ms CPU, GPU-free classification | Limited | <1ms CPU |
cis-lmu/glotlid (V3) | 4D | FastText-based (character n-grams) | 2,102 language labels Coverage (labels) | Sentence/paragraph | Low-resource & obscure dialect routing, preventing metadata leakage | Limited | <2ms CPU |
Microsoft Presidio In Use | Framework | Rule-based + spaCy NER + custom recognizers | 99%+ structured · ~80% names Accuracy (structured entities) | Unlimited (chunked internally) | Orchestration layer, rule-based PII (regex), plugging in any model above | Customizable | ~10–50ms CPU |
Why Segmento Sense
Built different. By design.
These aren't marketing checkboxes. They are deliberate architectural decisions that took 18 months to get right.
You Always Know Why — Not Just What
Most PII tools hand you a list of findings and leave you guessing. Sense shows you which model flagged each entity, which rule triggered it, and the exact text span that caused it. Transparency isn't a feature — it's the foundation.
You Control Precision vs. Recall
A single model is a single threshold. Our Consensus Engine lets you dial a confidence slider: low confidence flags aggressively (maximizes recall for regulated environments), high confidence flags conservatively (minimizes false positives for high-throughput pipelines). You choose the tradeoff.
Works Completely Offline — No Cloud Required
Banks, hospitals, and defense contractors cannot use SaaS tools. Sense runs all 18 models on your own private infrastructure with zero external API calls. Air-gap deployments are fully supported. Your data never crosses a network boundary it shouldn't.
Replace Real PII With Valid Synthetic Data
Redacting PII for test environments usually breaks data integrity. Sense replaces real PII with structurally valid synthetic data — SSNs that pass Luhn checks, IBANs with correct checksums, email addresses that match real domains. Your test data stays usable.
Generate SQL & Python Fix Scripts Automatically
Sense bridges the gap between the Security team that finds PII and the Data Engineering team that fixes it. One click generates a ready-to-run remediation script for the exact columns and tables flagged. No more "we found 3,000 PII fields" without a path forward.
The Industry Reality
Every major vendor sends your data to the cloud.
We built Sense to be the exception.
Whether it's a cloud platform charging per GB, an enterprise DLP tool requiring weeks of setup, or an AI-first SaaS that runs your sensitive documents through external model APIs — the industry's default is to move your data.
Segmento Sense runs all 18 models on your private infrastructure.Zero third-party AI access. Zero data egress. Zero compliance risk from vendor-side breaches. Your PII stays where it belongs — with you.
Ready to see the AI engine work for your data?
Upload a document. Watch all 18 models work in concert. See exactly which model flagged what entity, and why — in real time.