pii-protector 2.2.0__tar.gz → 2.2.2__tar.gz

{pii_protector-2.2.0 → pii_protector-2.2.2}/PKG-INFO

@@ -1,10 +1,10 @@
 Metadata-Version: 2.4
 Name: pii-protector
-Version: 2.2.0
+Version: 2.2.2
 Summary: Production-grade PII detection with multi-model ensemble
 Author-email: tensoryug <jainsatyam26@gmail.com>
 License-Expression: MIT
-Project-URL: Homepage, https://github.com/sjain26/pii-guard
+Project-URL: Homepage, https://github.com/sjain26/pii-detector
 Keywords: pii,privacy,nlp,detection,presidio,data-protection
 Classifier: Development Status :: 5 - Production/Stable
 Classifier: Intended Audience :: Developers
@@ -22,12 +22,12 @@ Requires-Dist: presidio-analyzer>=2.2; extra == "presidio"
 Requires-Dist: spacy>=3.7; extra == "presidio"
 Provides-Extra: transformers
 Requires-Dist: transformers>=4.40; extra == "transformers"
-Requires-Dist: torch>=2.0; extra == "transformers"
+Requires-Dist: optimum[onnxruntime]>=1.18; extra == "transformers"
 Provides-Extra: full
 Requires-Dist: presidio-analyzer>=2.2; extra == "full"
 Requires-Dist: spacy>=3.7; extra == "full"
 Requires-Dist: transformers>=4.40; extra == "full"
-Requires-Dist: torch>=2.0; extra == "full"
+Requires-Dist: optimum[onnxruntime]>=1.18; extra == "full"
 
 # pii-protector
 
@@ -40,9 +40,9 @@ Layer 1 — Regex                  always runs              ~0.3–3ms
     ↓  escalation score >= 3?
 Layer 2 — Presidio + spaCy NER   names / orgs / loc       ~5ms
     ↓  confidence low or conflict?
-Layer 3 — RoBERTa-large NER      high-accuracy names      ~20ms
+Layer 3 — NER Transformer        high-accuracy names      ~20ms
     +
-Layer 4 — piiranha-v1            structured PII           ~15ms
+Layer 4 — PII Model              structured PII           ~15ms
 ```
 
 Each layer decides whether the next one is needed. On clean structured text, only Layer 1 runs.
@@ -144,8 +144,8 @@ pii-detect --no-presidio --no-transformer --no-pii-model "card: 4111111111111111
 ```python
 detector = AdvancedPIIDetector(
     use_presidio=True,           # Layer 2: spaCy NER via Presidio
-    use_transformers=True,       # Layer 3: RoBERTa-large NER
-    use_pii_model=True,          # Layer 4: piiranha-v1
+    use_transformers=True,       # Layer 3: NER transformer
+    use_pii_model=True,          # Layer 4: PII model
     confidence_threshold=0.5,   # Early-stop threshold
     spacy_model="en_core_web_lg",
     transformer_model="jainsatyam26/pii-ner-onnx",
@@ -171,8 +171,8 @@ SPACY_DATA_PATH=/path/to/models
 | 4 | [`jainsatyam26/pii-detector-onnx`](https://huggingface.co/jainsatyam26/pii-detector-onnx) | PHONE, EMAIL, SSN, DOB, ADDRESS |
 
 ONNX-optimized versions:
-- [`jainsatyam26/pii-ner-onnx`](https://huggingface.co/jainsatyam26/pii-ner-onnx) — RoBERTa NER
-- [`jainsatyam26/pii-detector-onnx`](https://huggingface.co/jainsatyam26/pii-detector-onnx) — piiranha-v1
+- [`jainsatyam26/pii-ner-onnx`](https://huggingface.co/jainsatyam26/pii-ner-onnx) — NER model
+- [`jainsatyam26/pii-detector-onnx`](https://huggingface.co/jainsatyam26/pii-detector-onnx) — PII structured model
 
 ## License
 

{pii_protector-2.2.0 → pii_protector-2.2.2}/README.md

@@ -9,9 +9,9 @@ Layer 1 — Regex                  always runs              ~0.3–3ms
     ↓  escalation score >= 3?
 Layer 2 — Presidio + spaCy NER   names / orgs / loc       ~5ms
     ↓  confidence low or conflict?
-Layer 3 — RoBERTa-large NER      high-accuracy names      ~20ms
+Layer 3 — NER Transformer        high-accuracy names      ~20ms
     +
-Layer 4 — piiranha-v1            structured PII           ~15ms
+Layer 4 — PII Model              structured PII           ~15ms
 ```
 
 Each layer decides whether the next one is needed. On clean structured text, only Layer 1 runs.
@@ -113,8 +113,8 @@ pii-detect --no-presidio --no-transformer --no-pii-model "card: 4111111111111111
 ```python
 detector = AdvancedPIIDetector(
     use_presidio=True,           # Layer 2: spaCy NER via Presidio
-    use_transformers=True,       # Layer 3: RoBERTa-large NER
-    use_pii_model=True,          # Layer 4: piiranha-v1
+    use_transformers=True,       # Layer 3: NER transformer
+    use_pii_model=True,          # Layer 4: PII model
     confidence_threshold=0.5,   # Early-stop threshold
     spacy_model="en_core_web_lg",
     transformer_model="jainsatyam26/pii-ner-onnx",
@@ -140,8 +140,8 @@ SPACY_DATA_PATH=/path/to/models
 | 4 | [`jainsatyam26/pii-detector-onnx`](https://huggingface.co/jainsatyam26/pii-detector-onnx) | PHONE, EMAIL, SSN, DOB, ADDRESS |
 
 ONNX-optimized versions:
-- [`jainsatyam26/pii-ner-onnx`](https://huggingface.co/jainsatyam26/pii-ner-onnx) — RoBERTa NER
-- [`jainsatyam26/pii-detector-onnx`](https://huggingface.co/jainsatyam26/pii-detector-onnx) — piiranha-v1
+- [`jainsatyam26/pii-ner-onnx`](https://huggingface.co/jainsatyam26/pii-ner-onnx) — NER model
+- [`jainsatyam26/pii-detector-onnx`](https://huggingface.co/jainsatyam26/pii-detector-onnx) — PII structured model
 
 ## License
 

{pii_protector-2.2.0 → pii_protector-2.2.2}/pii_detector/__init__.py

@@ -7,5 +7,5 @@ Multi-model ensemble with pre-compiled patterns and type-aware voting
 from .detector import AdvancedPIIDetector, PIIEntity, PresidioCustomDetector
 from .config.settings import PII_CONFIG
 
-__version__ = "2.2.0"
+__version__ = "2.2.2"
 __all__ = ["AdvancedPIIDetector", "PIIEntity", "PresidioCustomDetector", "PII_CONFIG"]

{pii_protector-2.2.0 → pii_protector-2.2.2}/pii_detector/config/settings.py

@@ -4,8 +4,8 @@ Configuration settings for PII detection
 Latency profile (per-call on typical hardware):
   Regex only                          : ~0.3ms  (Layer 1 — structured PII, India PII, API secrets)
   Regex + Presidio                    : ~5ms    (Layer 2 — adds NAME/ORG/LOC via spaCy, CPU)
-  Regex + Presidio + RoBERTa-NER      : ~20ms   (Layer 3 — high-accuracy English NER, GPU)
-  + piiranha-v1                       : +15ms   (Layer 4 — structured PII: PHONE, EMAIL, DOB, SSN)
+  Regex + Presidio + NER Transformer   : ~20ms   (Layer 3 — high-accuracy English NER, GPU)
+  + PII Model                         : +15ms   (Layer 4 — structured PII: PHONE, EMAIL, DOB, SSN)
 
   Early stop fires for ~80% of texts → average well under 5ms.
 
@@ -14,7 +14,7 @@ Model selection rationale:
        - Higher NER accuracy than sm for NAME/ORG/LOC — worth the extra size
        - Used inside Presidio (Layer 2)
   ✅ jainsatyam26/pii-ner-onnx:
-       - Best-in-class English NER accuracy (RoBERTa-large fine-tuned on CoNLL-2003)
+       - Best-in-class English NER accuracy
        - Significantly better NAME detection vs multilingual models
        - Labels: PER → NAME, ORG → ORGANIZATION, LOC → LOCATION, MISC → _SKIP
   ✅ jainsatyam26/pii-detector-onnx:
@@ -27,11 +27,11 @@ Model selection rationale:
 PII_CONFIG = {
     # ── What to run ──────────────────────────────────────────────────
     # Presidio (spaCy NER): ON  → for NAME, ORG, LOCATION — CPU, fast
-    # Transformer (RoBERTa): ON  → high-accuracy English NER — GPU, conditional
-    # PII model (piiranha): ON  → structured PII specialist — always runs
+    # NER Transformer: ON  → high-accuracy English NER — GPU, conditional
+    # PII model: ON  → structured PII specialist — always runs
     'use_presidio': True,
-    'use_transformers': True,   # roberta-large NER enabled
-    'use_pii_model': True,      # piiranha enabled
+    'use_transformers': True,   # NER transformer enabled
+    'use_pii_model': True,      # PII model enabled
 
     # Model names
     'spacy_model': "en_core_web_lg",  # Large model — better NER accuracy than sm
@@ -98,7 +98,7 @@ TYPE_MAPPINGS = {
     'LAW': '_SKIP',
     'LANGUAGE': '_SKIP',
 
-    # Transformer types (Davlan/xlm-roberta-large-ner-hrl)
+    # Transformer types (multilingual NER)
     # NOTE: ORG -> ORGANIZATION and LOC -> LOCATION are already mapped in the spaCy section above
     'PER': 'NAME',          # Person names -> NAME
     'MISC': '_SKIP',        # Miscellaneous -> skip

{pii_protector-2.2.0 → pii_protector-2.2.2}/pii_detector/detector.py

@@ -82,7 +82,7 @@ def _empty_timing() -> Dict:
     """Return a zeroed timing dict for empty/skipped calls."""
     return {
         'regex_ms': 0.0, 'presidio_ms': 0.0, 'transformer_ms': 0.0,
-        'piiranha_ms': 0.0,
+        'pii_model_ms': 0.0,
         'escalation_check_ms': 0.0, 'postprocess_ms': 0.0, 'total_ms': 0.0,
         'layers': 'none', 'ner_escalation_score': 0,
         'entities_found': 0, 'text_length': 0,
@@ -148,7 +148,7 @@ _COMMON_NON_NAME_WORDS = frozenset({
 
 # Escalation score thresholds
 _NER_ESCALATION_THRESHOLD = 3
-# Transformer (Davlan multilingual) runs only when Presidio NER is uncertain.
+# NER transformer runs only when Presidio NER is uncertain.
 # Threshold=3 means: run transformer only if name confidence is low or conflicting.
 # 0 = always run (wasteful), 3 = smart conditional (saves ~9ms on clear cases).
 _TRANSFORMER_ESCALATION_THRESHOLD = 3
@@ -169,7 +169,7 @@ class PIIEntity:
         return hash((self.text, self.entity_type, self.start, self.end))
 
 
-# Person-context keywords — when present, a piiranha USERNAME is likely a NAME.
+# Person-context keywords — when present, a pii_model USERNAME is likely a NAME.
 _PERSON_CONTEXT_KW = frozenset({
     'appointment', 'scheduling', 'schedule', 'scheduled',
     'patient', 'customer', 'client', 'employee', 'user',
@@ -178,7 +178,7 @@ _PERSON_CONTEXT_KW = frozenset({
     'meeting', 'visit', 'booking', 'reservation',
 })
 
-# piiranha-v1 actual label → our standard type (labels verified from model output).
+# PII model actual label → our standard type (labels verified from model output).
 _PIIRANHA_MAP = {
     # Name variants
     'GIVENNAME':          'NAME',
@@ -215,7 +215,7 @@ _PIIRANHA_MAP = {
     'ID_CARD':            '_SKIP',
     # Auth
     'USERNAME':           'USERNAME',
-    'PASSWORD':           '_SKIP',  # NOSONAR - piiranha label key, not a credential
+    'PASSWORD':           '_SKIP',  # NOSONAR - pii_model label key, not a credential
     # Location (skip — too noisy)
     'CITY':               '_SKIP',
     'ZIPCODE':            '_SKIP',
@@ -305,23 +305,42 @@ class AdvancedPIIDetector:
         self.type_mappings = TYPE_MAPPINGS
 
         self.presidio = None
-        self.ner_pipeline = None   # RoBERTa-large English NER: NAME / ORG / LOC
-        self.pii_pipeline = None   # piiranha-v1: structured PII (PHONE, EMAIL, SSN, DOB etc.)
+        self.ner_pipeline = None   # NER transformer: NAME / ORG / LOC
+        self.pii_pipeline = None   # PII model: structured PII (PHONE, EMAIL, SSN, DOB etc.)
         self._ner_sources_count = 0
         self.last_timing: Dict = _empty_timing()
 
         if self.use_presidio:
             self._init_presidio()
         if self.use_transformers:
-            self._init_transformers()  # RoBERTa-large English NER (NAME/ORG/LOC)
+            self._init_transformers()  # NER transformer (NAME/ORG/LOC)
         if self.use_pii_model:
-            self._init_pii_transformer()  # piiranha-v1 structured PII
+            self._init_pii_transformer()  # PII model structured PII
+
+        if self.use_presidio and not self.presidio:
+            warnings.warn(
+                "Presidio (Layer 2) unavailable — NAME/ORG/LOCATION detection disabled. "
+                "Install: pip install 'pii-protector[presidio]' then: python -m spacy download en_core_web_lg",
+                UserWarning, stacklevel=2,
+            )
+        if self.use_transformers and not self.ner_pipeline:
+            warnings.warn(
+                "NER transformer (Layer 3) unavailable — high-accuracy name detection disabled. "
+                "Install: pip install 'optimum[onnxruntime]>=1.18' 'transformers>=4.40'",
+                UserWarning, stacklevel=2,
+            )
+        if self.use_pii_model and not self.pii_pipeline:
+            warnings.warn(
+                "PII model (Layer 4) unavailable — PHONE/EMAIL/DOB/SSN transformer detection disabled. "
+                "Install: pip install 'optimum[onnxruntime]>=1.18' 'transformers>=4.40'",
+                UserWarning, stacklevel=2,
+            )
 
         logger.info(
             f"PII Detector ready | mode=SEQUENTIAL_LAYERED "
             f"| presidio={'ON' if self.presidio else 'OFF'} "
             f"| ner={'ON' if self.ner_pipeline else 'OFF'} "
-            f"| piiranha={'ON' if self.pii_pipeline else 'OFF'} "
+            f"| pii_model={'ON' if self.pii_pipeline else 'OFF'} "
             f"| ner_sources={self._ner_sources_count}"
         )
 
@@ -363,61 +382,43 @@ class AdvancedPIIDetector:
             logger.warning(f"Presidio unavailable: {e}")
 
     def _init_transformers(self):
-        """RoBERTa-large English NER — high-accuracy NAME/ORG/LOC."""
+        """NER transformer — high-accuracy NAME/ORG/LOC (ONNX Runtime)."""
         try:
-            from transformers import pipeline
-            try:
-                with _suppress_stdout():
-                    self.ner_pipeline = pipeline(
-                        "ner",
-                        model=self.transformer_model_name,
-                        aggregation_strategy="first",
-                        device=0,
-                        model_kwargs={"dtype": "float16"},
-                    )
-                self.ner_pipeline("warmup")  # GPU kernel compile
-                logger.info(f"NER ready on GPU ({self.transformer_model_name})")
-            except Exception:
-                with _suppress_stdout():
-                    self.ner_pipeline = pipeline(
-                        "ner",
-                        model=self.transformer_model_name,
-                        aggregation_strategy="first",
-                        device=-1,
-                    )
-                self.ner_pipeline("warmup")
-                logger.info(f"NER ready on CPU ({self.transformer_model_name})")
+            from optimum.onnxruntime import ORTModelForTokenClassification
+            from transformers import AutoTokenizer, pipeline
+            with _suppress_stdout():
+                model = ORTModelForTokenClassification.from_pretrained(self.transformer_model_name)
+                tokenizer = AutoTokenizer.from_pretrained(self.transformer_model_name)
+            self.ner_pipeline = pipeline(
+                "ner",
+                model=model,
+                tokenizer=tokenizer,
+                aggregation_strategy="first",
+            )
+            self.ner_pipeline("warmup")
             self._ner_sources_count += 1
+            logger.info(f"NER ready ({self.transformer_model_name})")
         except Exception as e:
             logger.warning(f"NER transformer unavailable: {e}")
 
     def _init_pii_transformer(self):
-        """piiranha-v1 structured PII detection — PHONE, EMAIL, SSN, DOB, ADDRESS etc."""
+        """PII model structured PII detection — PHONE, EMAIL, SSN, DOB, ADDRESS etc. (ONNX Runtime)."""
         try:
-            from transformers import pipeline
-            try:
-                with _suppress_stdout():
-                    self.pii_pipeline = pipeline(
-                        "token-classification",
-                        model=self.pii_model_name,
-                        aggregation_strategy="first",
-                        device=0,
-                        model_kwargs={"dtype": "float16"},
-                    )
-                self.pii_pipeline("warmup")  # GPU kernel compile
-                logger.info(f"piiranha ready on GPU ({self.pii_model_name})")
-            except Exception:
-                with _suppress_stdout():
-                    self.pii_pipeline = pipeline(
-                        "token-classification",
-                        model=self.pii_model_name,
-                        aggregation_strategy="first",
-                        device=-1,
-                    )
-                self.pii_pipeline("warmup")
-                logger.info(f"piiranha ready on CPU ({self.pii_model_name})")
+            from optimum.onnxruntime import ORTModelForTokenClassification
+            from transformers import AutoTokenizer, pipeline
+            with _suppress_stdout():
+                model = ORTModelForTokenClassification.from_pretrained(self.pii_model_name)
+                tokenizer = AutoTokenizer.from_pretrained(self.pii_model_name)
+            self.pii_pipeline = pipeline(
+                "token-classification",
+                model=model,
+                tokenizer=tokenizer,
+                aggregation_strategy="first",
+            )
+            self.pii_pipeline("warmup")
+            logger.info(f"PII model ready ({self.pii_model_name})")
         except Exception as e:
-            logger.warning(f"piiranha unavailable: {e}")
+            logger.warning(f"PII model unavailable: {e}")
 
     # ------------------------------------------------------------------
     # Layer detection methods
@@ -540,8 +541,8 @@ class AdvancedPIIDetector:
         clean = word.strip()
         return clean.isalpha() and len(clean) >= 3
 
-    def _piiranha_results_to_entities(self, results, offset, overlap, person_ctx):
-        """Map piiranha pipeline results (one chunk) into PIIEntity objects."""
+    def _pii_model_results_to_entities(self, results, offset, overlap, person_ctx):
+        """Map pii_model pipeline results (one chunk) into PIIEntity objects."""
         out = []
         for r in results:
             if offset > 0 and r['start'] < overlap:
@@ -551,23 +552,23 @@ class AdvancedPIIDetector:
             if entity_type == '_SKIP':
                 continue
             word = r['word']
-            # piiranha often tags standalone person names as USERNAME in free text;
+            # PII model often tags standalone person names as USERNAME in free text;
             # promote to NAME when person context is present and the word looks like a name.
             if entity_type == 'USERNAME' and person_ctx and self._looks_like_name(word):
                 entity_type = 'NAME'
-                logger.debug(f"[piiranha] Promoted USERNAME→NAME for '{word}' (person context)")
+                logger.debug(f"[pii_model] Promoted USERNAME→NAME for '{word}' (person context)")
             out.append(PIIEntity(
                 text=word,
                 entity_type=entity_type,
                 start=int(r['start']) + offset,
                 end=int(r['end']) + offset,
                 score=float(r['score']),
-                source="piiranha",
+                source="pii_model",
             ))
         return out
 
     def _detect_pii_transformer(self, text: str) -> List[PIIEntity]:
-        """piiranha-v1 specialized PII detection with chunking (see _PIIRANHA_MAP)."""
+        """PII model specialized PII detection with chunking (see _PIIRANHA_MAP)."""
         if not self.pii_pipeline:
             return []
 
@@ -584,11 +585,11 @@ class AdvancedPIIDetector:
                 for chunk_results, chunk_start in zip(
                     self.pii_pipeline(chunks, batch_size=batch_size), offsets
                 ):
-                    entities.extend(self._piiranha_results_to_entities(chunk_results, chunk_start, overlap, person_ctx))
+                    entities.extend(self._pii_model_results_to_entities(chunk_results, chunk_start, overlap, person_ctx))
             else:
-                entities.extend(self._piiranha_results_to_entities(self.pii_pipeline(text), 0, overlap, person_ctx))
+                entities.extend(self._pii_model_results_to_entities(self.pii_pipeline(text), 0, overlap, person_ctx))
         except Exception as e:
-            logger.warning(f"PII transformer (piiranha) detection failed: {e}")
+            logger.warning(f"PII transformer detection failed: {e}")
 
         return entities
 
@@ -640,7 +641,7 @@ class AdvancedPIIDetector:
 
         When non-Latin Unicode is present the text is in a language that static
         English/Latin keyword patterns cannot cover.  We must escalate to the
-        XLM-RoBERTa multilingual transformer so it can detect names/orgs/locations.
+        multilingual NER transformer so it can detect names/orgs/locations.
         """
         return bool(re.search(
             r'[\u0600-\u06FF'   # Arabic
@@ -709,7 +710,7 @@ class AdvancedPIIDetector:
             score += 3
         if has_ner_kw:
             score += 3
-        if has_non_latin:          # multilingual text needs XLM-RoBERTa NER
+        if has_non_latin:          # multilingual text needs multilingual NER
             score += 3
 
         score += self._regex_confidence_score(
@@ -759,7 +760,7 @@ class AdvancedPIIDetector:
         Signals:
           +3  NER entity with confidence < 0.55 (very uncertain)
           +3  Non-Latin Unicode present — Presidio/spaCy is English-only,
-              XLM-RoBERTa transformer is the only model that handles
+              multilingual NER transformer is the only model that handles
               Arabic / CJK / Devanagari / Cyrillic NER correctly.
           +2  NER entity with confidence 0.55 – 0.70 (uncertain)
           +2  Overlapping entities of different types (conflict)
@@ -774,11 +775,11 @@ class AdvancedPIIDetector:
         ner_types = {'NAME', 'PERSON', 'ORGANIZATION', 'LOCATION'}
         ner_ents = [e for e in current_entities if e.entity_type in ner_types]
 
-        # Non-Latin Unicode → spaCy is English-only; escalate to XLM-RoBERTa.
+        # Non-Latin Unicode: escalate to multilingual NER.
         if self._has_non_latin_unicode(text):
             score += 3
 
-        # RoBERTa-large is markedly more accurate than spaCy NER for person names.
+        # NER transformer is markedly more accurate than spaCy NER for person names.
         if any(e.entity_type in {'NAME', 'PERSON'} for e in ner_ents):
             score += 3
 
@@ -831,7 +832,7 @@ class AdvancedPIIDetector:
             return False
 
         # Never early-stop for non-Latin Unicode (Arabic, CJK, Devanagari etc.)
-        # Static keyword patterns are Latin-only; XLM-RoBERTa must run.
+        # Static keyword patterns are Latin-only; multilingual NER must run.
         if self._has_non_latin_unicode(text):
             return False
 
@@ -1037,28 +1038,28 @@ class AdvancedPIIDetector:
     # ------------------------------------------------------------------
 
     def _run_ner_layers(self, text, all_entities, layers_used, timing):
-        """Run Presidio + RoBERTa + piiranha layers, mutating all_entities/layers_used/timing."""
+        """Run Presidio + NER + PII model layers, mutating all_entities/layers_used/timing."""
         if self.presidio:
             t0 = time.perf_counter()
             all_entities.extend(self._detect_presidio(text))
             timing['presidio_ms'] = _elapsed_ms(t0)
             layers_used.append('presidio')
 
-        # Layer 3 (RoBERTa-large) runs only when Presidio NER is uncertain.
+        # Layer 3 (NER transformer) runs only when Presidio NER is uncertain.
         if self.ner_pipeline:
             transformer_score = self._calc_transformer_escalation_score(text, list(all_entities))
             if transformer_score >= _TRANSFORMER_ESCALATION_THRESHOLD:
                 t0 = time.perf_counter()
                 all_entities.extend(self._detect_transformers(text))
                 timing['transformer_ms'] = _elapsed_ms(t0)
-                layers_used.append('roberta')
+                layers_used.append('ner')
 
-        # Layer 4 (piiranha-v1) catches informal PHONE/EMAIL/SSN/DOB that regex misses.
+        # Layer 4 (PII model) catches informal PHONE/EMAIL/SSN/DOB that regex misses.
         if self.pii_pipeline:
             t0 = time.perf_counter()
             all_entities.extend(self._detect_pii_transformer(text))
-            timing['piiranha_ms'] = _elapsed_ms(t0)
-            layers_used.append('piiranha')
+            timing['pii_model_ms'] = _elapsed_ms(t0)
+            layers_used.append('pii_model')
 
     @staticmethod
     def _entities_to_dicts(entities):
@@ -1083,8 +1084,8 @@ class AdvancedPIIDetector:
           2. Early stop check         -> skip NER if high confidence + no name signals
           3. Escalation score         -> if >= 3, run NER layers
           4. Layer 2 (Presidio NER)   -> spaCy NAME, ORG, LOCATION
-          5. Layer 3 (roberta-large)  -> high-accuracy NAME detection
-          6. Layer 4 (piiranha-v1)    -> structured PII: PHONE, EMAIL, SSN, DOB etc.
+          5. Layer 3 (NER transformer) -> high-accuracy NAME detection
+          6. Layer 4 (PII model)      -> structured PII: PHONE, EMAIL, SSN, DOB etc.
           7. Ensemble vote -> conflict resolve -> filter -> deduplicate
 
         Returns list of entity dicts.
@@ -1114,7 +1115,7 @@ class AdvancedPIIDetector:
         layers_used = ['regex']
         timing['presidio_ms'] = 0.0
         timing['transformer_ms'] = 0.0
-        timing['piiranha_ms'] = 0.0
+        timing['pii_model_ms'] = 0.0
         timing['early_stop'] = False
         
         # ── Early stopping check (if high confidence) ─────────────────
@@ -1160,8 +1161,8 @@ class AdvancedPIIDetector:
             f"[PII TIMER] total={timing['total_ms']:.1f}ms | "
             f"regex={timing['regex_ms']:.1f}ms | "
             f"presidio={timing['presidio_ms']:.1f}ms | "
-            f"roberta={timing['transformer_ms']:.1f}ms | "
-            f"piiranha={timing['piiranha_ms']:.1f}ms | "
+            f"ner={timing['transformer_ms']:.1f}ms | "
+            f"pii_model={timing['pii_model_ms']:.1f}ms | "
             f"post={timing['postprocess_ms']:.1f}ms"
         )
 
@@ -1242,6 +1243,6 @@ class PresidioCustomDetector(AdvancedPIIDetector):
             use_presidio=PII_CONFIG.get('use_presidio', True),
             use_transformers=PII_CONFIG.get('use_transformers', True),
             transformer_model=PII_CONFIG.get('transformer_model',
-                                              'Davlan/bert-base-multilingual-cased-ner-hrl'),
+                                              'jainsatyam26/pii-ner-onnx'),
         )
         self.name = "Advanced_PII_Detector"

{pii_protector-2.2.0 → pii_protector-2.2.2}/pii_detector/patterns/core_pii.py

@@ -169,7 +169,7 @@ CORE_PII_PATTERNS = {
     # ── NAME ───────────────────────────────────────────────────────────
     # Regex only covers STRUCTURED label patterns (exact field names).
     # Free-text names (threat verbs, sentences) are handled by
-    # Presidio/spaCy NER (Layer 2) + RoBERTa transformer (Layer 3).
+    # Presidio/spaCy NER (Layer 2) + NER transformer (Layer 3).
     'NAME': [
         # Explicit form/document field labels — unambiguous
         # "Full Name: John Smith", "Employee Name: Sneha Patel", "Claimant Name: Ahmed Khan"

{pii_protector-2.2.0 → pii_protector-2.2.2}/pii_detector/utils/helpers.py

@@ -111,8 +111,8 @@ def deduplicate_entities(entities: List[Dict], overlap_threshold: float = 0.5) -
       1. Score ke basis pe descending sort karo (highest score pehle process ho)
       2. Har entity ke liye check karo — kya kisi already-kept entity se overlap hai?
          - Same span + SAME type   → duplicate (same detector found it twice)
-         - Same span + DIFF type   → cross-category conflict (e.g. piiranha EMAIL vs
-                                     regex EMAIL, or piiranha PHONE vs regex PHONE) —
+         - Same span + DIFF type   → cross-category conflict (e.g. pii_model EMAIL vs
+                                     regex EMAIL, or pii_model PHONE vs regex PHONE) —
                                      keep only highest-score, drop the rest
       3. Agar overlap hai → discard (already-kept wala higher score ka hai)
       4. Final result position ke basis pe sort karo

{pii_protector-2.2.0 → pii_protector-2.2.2}/pii_protector.egg-info/PKG-INFO

@@ -1,10 +1,10 @@
 Metadata-Version: 2.4
 Name: pii-protector
-Version: 2.2.0
+Version: 2.2.2
 Summary: Production-grade PII detection with multi-model ensemble
 Author-email: tensoryug <jainsatyam26@gmail.com>
 License-Expression: MIT
-Project-URL: Homepage, https://github.com/sjain26/pii-guard
+Project-URL: Homepage, https://github.com/sjain26/pii-detector
 Keywords: pii,privacy,nlp,detection,presidio,data-protection
 Classifier: Development Status :: 5 - Production/Stable
 Classifier: Intended Audience :: Developers
@@ -22,12 +22,12 @@ Requires-Dist: presidio-analyzer>=2.2; extra == "presidio"
 Requires-Dist: spacy>=3.7; extra == "presidio"
 Provides-Extra: transformers
 Requires-Dist: transformers>=4.40; extra == "transformers"
-Requires-Dist: torch>=2.0; extra == "transformers"
+Requires-Dist: optimum[onnxruntime]>=1.18; extra == "transformers"
 Provides-Extra: full
 Requires-Dist: presidio-analyzer>=2.2; extra == "full"
 Requires-Dist: spacy>=3.7; extra == "full"
 Requires-Dist: transformers>=4.40; extra == "full"
-Requires-Dist: torch>=2.0; extra == "full"
+Requires-Dist: optimum[onnxruntime]>=1.18; extra == "full"
 
 # pii-protector
 
@@ -40,9 +40,9 @@ Layer 1 — Regex                  always runs              ~0.3–3ms
     ↓  escalation score >= 3?
 Layer 2 — Presidio + spaCy NER   names / orgs / loc       ~5ms
     ↓  confidence low or conflict?
-Layer 3 — RoBERTa-large NER      high-accuracy names      ~20ms
+Layer 3 — NER Transformer        high-accuracy names      ~20ms
     +
-Layer 4 — piiranha-v1            structured PII           ~15ms
+Layer 4 — PII Model              structured PII           ~15ms
 ```
 
 Each layer decides whether the next one is needed. On clean structured text, only Layer 1 runs.
@@ -144,8 +144,8 @@ pii-detect --no-presidio --no-transformer --no-pii-model "card: 4111111111111111
 ```python
 detector = AdvancedPIIDetector(
     use_presidio=True,           # Layer 2: spaCy NER via Presidio
-    use_transformers=True,       # Layer 3: RoBERTa-large NER
-    use_pii_model=True,          # Layer 4: piiranha-v1
+    use_transformers=True,       # Layer 3: NER transformer
+    use_pii_model=True,          # Layer 4: PII model
     confidence_threshold=0.5,   # Early-stop threshold
     spacy_model="en_core_web_lg",
     transformer_model="jainsatyam26/pii-ner-onnx",
@@ -171,8 +171,8 @@ SPACY_DATA_PATH=/path/to/models
 | 4 | [`jainsatyam26/pii-detector-onnx`](https://huggingface.co/jainsatyam26/pii-detector-onnx) | PHONE, EMAIL, SSN, DOB, ADDRESS |
 
 ONNX-optimized versions:
-- [`jainsatyam26/pii-ner-onnx`](https://huggingface.co/jainsatyam26/pii-ner-onnx) — RoBERTa NER
-- [`jainsatyam26/pii-detector-onnx`](https://huggingface.co/jainsatyam26/pii-detector-onnx) — piiranha-v1
+- [`jainsatyam26/pii-ner-onnx`](https://huggingface.co/jainsatyam26/pii-ner-onnx) — NER model
+- [`jainsatyam26/pii-detector-onnx`](https://huggingface.co/jainsatyam26/pii-detector-onnx) — PII structured model
 
 ## License
 

{pii_protector-2.2.0 → pii_protector-2.2.2}/pii_protector.egg-info/requires.txt

@@ -3,7 +3,7 @@
 presidio-analyzer>=2.2
 spacy>=3.7
 transformers>=4.40
-torch>=2.0
+optimum[onnxruntime]>=1.18
 
 [presidio]
 presidio-analyzer>=2.2
@@ -11,4 +11,4 @@ spacy>=3.7
 
 [transformers]
 transformers>=4.40
-torch>=2.0
+optimum[onnxruntime]>=1.18

{pii_protector-2.2.0 → pii_protector-2.2.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "pii-protector"
-version = "2.2.0"
+version = "2.2.2"
 description = "Production-grade PII detection with multi-model ensemble"
 readme = "README.md"
 license = "MIT"
@@ -35,24 +35,24 @@ presidio = [
     "presidio-analyzer>=2.2",
     "spacy>=3.7",
 ]
-# Layer 3 & 4: Transformer-based NER (RoBERTa) + piiranha-v1
+# Layer 3 & 4: NER transformer + PII model (ONNX Runtime)
 transformers = [
     "transformers>=4.40",
-    "torch>=2.0",
+    "optimum[onnxruntime]>=1.18",
 ]
 # Install everything
 full = [
     "presidio-analyzer>=2.2",
     "spacy>=3.7",
     "transformers>=4.40",
-    "torch>=2.0",
+    "optimum[onnxruntime]>=1.18",
 ]
 
 [project.scripts]
 pii-detect = "pii_detector.cli:main"
 
 [project.urls]
-Homepage = "https://github.com/sjain26/pii-guard"
+Homepage = "https://github.com/sjain26/pii-detector"
 
 [tool.setuptools.packages.find]
 include = ["pii_detector*"]