Update PII detector model card

.gitattributes

@@ -35,4 +35,3 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 leaderboard.png filter=lfs diff=lfs merge=lfs -text
 leaderboard_18lang.png filter=lfs diff=lfs merge=lfs -text
-per_language_internal.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -53,13 +53,16 @@ base_model:
 
 # LFM2.5-Encoder-350-PII-Detector
 
-A full fine-tune of [LFM2.5-Encoder-350M](https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M) with a token-classification head, covering **40 PII types** across **16 languages**. 
+A full fine-tune of [LFM2.5-Encoder-350M](https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M) 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-timem**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.
+
+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.
 
 Find more details about our encoders in our [blog post](https://www.liquid.ai/blog/lfm2-5-encoders).
 
 > [!NOTE]
 > 💻 **Demos**: Try this fine-tuned model running in a CPU-only Hugging Face space:
-> **[PII detection](https://huggingface.co/spaces/LiquidAI/pii-detection)** — spot and remove 40 kinds of personal information across 16 languages.
+> [PII detection](https://huggingface.co/spaces/LiquidAI/pii-detection)** — spot and remove 40 kinds of personal information across 16 languages.
 
 ## Entity types (40 PII types across 11 domains)
 
@@ -79,25 +82,22 @@ Find more details about our encoders in our [blog post](https://www.liquid.ai/bl
 
 ## Benchmarks (18-locale-filtered, partial-F1, hybrid decode)
 
-| Benchmark | **this model** | detection tier | SauerkrautLM GLiNER | openai/privacy-filter | Piiranha-v1 | OpenMed privacy-filter | regex + validators |
-|---|---|---|---|---|---|---|---|
-| SPY | **0.428** | 0.509 | 0.280 | 0.264 | 0.232 | 0.226 | 0.358 |
-| Gretel | **0.880** | 0.885 | 0.663 | 0.458 | 0.553 | 0.770 | 0.337 |
-| TAB | **0.867** | 0.888 | 0.685 | 0.543 | 0.262 | 0.672 | 0.000 |
-| ai4privacy | **0.715** | 0.774 | 0.488 | 0.394 | 0.946 | 0.432 | 0.195 |
-| Nemotron | **0.855** | 0.863 | 0.639 | 0.572 | 0.658 | 0.918 | 0.335 |
-| MAPA | **0.236** | 0.267 | 0.416 | 0.288 | 0.228 | 0.164 | 0.000 |
+| 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)
 
-- **Best on every benchmark except MAPA**, whose idiosyncratic date-as-`date_of_birth` labeling
-  convention penalises correctly-typed predictions. Only two external scores land higher
-  anywhere — Piiranha-v1's 0.946 on ai4privacy and OpenMed's 0.918 on Nemotron — and both are
-  in-distribution: each model trains on that exact corpus, as did our own encoder's pretraining
-  on those same two.
-- **Detection tier** is the same model and the same predictions, scored with the type label
-  ignored — did it find the PII span at all, which is the metric that matters for redaction. The
-  gap to exact-type is the type-confusion rate, e.g. SPY 0.428 → 0.509.
+- **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