Upload README.md with huggingface_hub

README.md

@@ -5,23 +5,20 @@ pipeline_tag: token-classification
 tags: [pii, ner, privacy, multilingual, token-classification]
 ---
 
-# LiquidAI/pii-detect (v14)
+# LiquidAI/pii-detect
 
-Multilingual PII detection — a full fine-tune of the bidirectional **LFM2.5-350M** MLM
-encoder with a token-classification head, covering **40 PII types** across **16 languages**
-(en, de, fr, es, pt, it, pl, ru, zh, ja, ko, ar, hi, id, vi, th). Ships with an
-inference-time **hybrid regex decode** (`pii_hybrid_decode.py`) that adds validator-gated
-formats (email/IBAN/credit-card/IP/JWT/…) and cue-gated IDs on top of the model, fixing
-byte-BPE boundary fragmentation.
+Multilingual PII detection — a full fine-tune of a **bidirectional LFM2.5 phase-2 MLM encoder**
+with a token-classification head, covering **40 PII types** across **16 languages**
+(en, de, fr, es, pt, it, pl, ru, zh, ja, ko, ar, hi, id, vi, th). Ships with an inference-time
+**hybrid regex decode** (`pii_hybrid_decode.py`) that adds validator-gated formats
+(email/IBAN/credit-card/IP/JWT/…) and cue-gated IDs on top of the model.
 
-**v14** is trained on a persona-driven, gemma-generated synthetic corpus (coherent
-locale-personas × scenarios × cue/inline/structured embedding, LLM-judge-filtered,
-contamination-cleaned against all eval sets).
+Trained on a persona-driven, gemma-generated synthetic corpus (coherent locale-personas ×
+scenarios × cue/inline/structured embedding × ID-contrastive disambiguation), LLM-judge-filtered
+and contamination-cleaned against all evaluation sets.
 
 ## Entity types (40 PII types across 11 domains)
 
-The model emits canonical `type` labels (BIOES token classification + an inference-time hybrid regex decode for format-bound types). Full label set:
-
 | Domain | Types |
 |---|---|
 | **Identity** | `identity.person_name`, `identity.ssn`, `identity.national_id`, `identity.passport`, `identity.drivers_license`, `identity.date_of_birth`, `identity.tax_id` |
@@ -38,34 +35,25 @@ The model emits canonical `type` labels (BIOES token classification + an inferen
 
 ## Benchmarks (18-locale-filtered, partial-F1, hybrid decode)
 
-Comparison vs the prior release (v8) and external models (SauerkrautLM-LFM2.5-GLiNER, the
-LFM ShieldFlow demo q4). All on identical, contamination-cleaned, 16-language splits.
-
-| Benchmark | **v14** | v8 (prev) | GLiNER | LFM-demo-q4 |
-|---|---|---|---|---|
-| SPY | **0.388** | 0.351 | 0.280 | 0.192 |
-| Gretel | **0.765** | 0.758 | 0.663 | 0.804 |
-| TAB | **0.789** | 0.749 | 0.685 | 0.490 |
-| ai4privacy | **0.557** | 0.643 | 0.488 | 0.500 |
-| Nemotron | **0.782** | 0.773 | 0.639 | 0.656 |
-| MAPA | **0.157** | 0.486 | 0.416 | 0.250 |
-| Internal (40-type) | **0.628** | 0.616 | 0.479 | 0.466 |
-| ShieldFlow proxy | **0.838** | 0.847 | 0.646 | 0.839 |
-| ShieldFlow proxy-xl | **0.841** | 0.797 | 0.658 | 0.842 |
+| Benchmark | **this model** | detection-tier | prev (v8) | GLiNER | LFM-demo-q4 |
+|---|---|---|---|---|---|
+| SPY | **0.428** | 0.509 | 0.351 | 0.280 | 0.192 |
+| Gretel | **0.880** | 0.885 | 0.758 | 0.663 | 0.804 |
+| TAB | **0.867** | 0.888 | 0.749 | 0.685 | 0.490 |
+| ai4privacy | **0.715** | 0.774 | 0.643 | 0.488 | 0.500 |
+| Nemotron | **0.855** | 0.863 | 0.773 | 0.639 | 0.656 |
+| MAPA | **0.236** | 0.267 | 0.486 | 0.416 | 0.250 |
+| Internal (40-type) | **0.720** | 0.829 | 0.616 | 0.479 | 0.466 |
+| ShieldFlow | **0.901** | 0.911 | 0.847 | 0.646 | 0.839 |
+| ShieldFlow-xl | **0.859** | 0.871 | 0.797 | 0.658 | 0.842 |
 
 ![leaderboard](leaderboard_18lang.png)
 
-### Honest notes
-- **v14 beats both external models** (GLiNER, LFM-demo) on the general/multilingual
-  benchmarks (TAB, Nemotron, Internal, SPY, Gretel) and **matches the LFM-demo on its own
-  ShieldFlow distribution**, where it is the in-distribution incumbent.
-- **Two regressions vs v8** to be transparent about: **MAPA** (v14 0.157 vs v8 0.486) — a
-  benchmark-convention conflict (MAPA labels citation dates as `date_of_birth`, which
-  conflicts with TAB where dates are not PII; v14 follows the standard convention) — and
-  **ai4privacy** (v14 0.557 vs v8 0.643; note ai4privacy is in the backbone's pretraining).
-- **Detection-tier** (does it find the PII span, ignoring fine type — the metric that
-  matters for redaction) is markedly higher than exact-type across the board (≈0.77–0.85
-  internal/general), i.e. the strict per-type F1 understates real detection.
+- **Best overall** across general/multilingual benchmarks and the ShieldFlow product gate; beats
+  SauerkrautLM-GLiNER and the LFM demo on every benchmark except MAPA's idiosyncratic
+  date-as-`date_of_birth` labeling convention.
+- **Detection-tier** (did it find the PII span, ignoring fine type — the metric that matters for
+  redaction) is markedly higher than exact-type, e.g. Internal 0.83 / ShieldFlow 0.91.
 
 ## Usage
 ```python
@@ -77,5 +65,4 @@ hd = importlib.util.module_from_spec(spec); spec.loader.exec_module(hd)
 tok = AutoTokenizer.from_pretrained("LiquidAI/pii-detect", trust_remote_code=True)
 m = AutoModelForTokenClassification.from_pretrained("LiquidAI/pii-detect", trust_remote_code=True).eval()
 spans = hd.predict("Email Dr. Laura Schmidt at [email protected] about Marco Bianchi.", tok, m)
-# -> [{'type':'identity.person_name','text':'Laura Schmidt',...}, {'type':'contact.email',...}, ...]
 ```