Add OSSRedact PII NER v11r9c weights + model card (bilingual FR/EN, MIT)

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+banner.png filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,141 @@
+---
+license: mit
+language:
+  - fr
+  - en
+library_name: transformers.js
+pipeline_tag: token-classification
+base_model: FacebookAI/xlm-roberta-base
+tags:
+  - pii
+  - redaction
+  - privacy
+  - ner
+  - token-classification
+  - french
+  - quebec
+  - onnx
+  - int8
+  - gdpr
+  - law-25
+---
+
+# OSSRedact PII NER -- xlm-roberta-base (CPU / in-browser tier)
+
+![OSSRedact PII (base) -- in-browser bilingual FR/EN PII + secrets detection](banner.png)
+
+**A bilingual (French-Québec + English) PII / secrets token-classifier**, dynamic-INT8 ONNX -- the always-on
+detection model behind [OSSRedact](https://github.com/ZenSystemAI/OSSRedact), a local privacy gateway that
+redacts private data before it reaches a cloud LLM and rehydrates it on the reply.
+
+Shipping revision: **`v11r9c`** (carries the cumulative organization/address augmentation; address recall is now
+~0.93 on this tier). The higher-capacity tier is
+[`ZenSystemAI/ossredact-pii-large`](https://huggingface.co/ZenSystemAI/ossredact-pii-large).
+
+![OSSRedact vs Microsoft Presidio on Québec FR/EN PII](fig5_vs_presidio.png)
+
+![Recall by tier](fig1_recall_by_tier.png)
+
+*OSSRedact recall vs Microsoft Presidio (historical v6/v7 sets), and recall by tier. The base model nearly
+matches the large model on recall (trailing by ~1 point on overall recall) at roughly 4x lower latency -- which
+is why it is the always-on tier.*
+
+## What it is
+
+A `xlm-roberta-base` token classifier fine-tuned to tag **20 PII / secret entity types** in realistic French-
+Québec and English documents, exported to **dynamic-INT8 ONNX (~277 MB)** for CPU (onnxruntime) and **in-browser**
+(transformers.js / onnxruntime-web) inference. The carded numbers below are the fp32 `v11r9c` reference; the
+shipped artifact is the **per-channel dynamic INT8** export (WASM-native). v11r9c's org/address augmentation
+sharpened the boundaries, so the INT8 lands at pii_argmax 0.967 (cosine 0.997, faithful) -- the parity bar is
+0.965 for this reason: ~62% of the token-flips are on floor-protected types the deterministic Tier-0 layer
+redacts regardless of the model, and person (the highest-frequency no-floor type) is barely affected.
+It is the always-on detection tier of OSSRedact; in production it
+runs **inside** the OSSRedact gateway alongside a deterministic Tier-0 floor. Detection runs **locally** -- no
+call leaves the machine; in the browser, the document never leaves the page.
+
+The bilingual Québec-French focus is the differentiator: general English-first PII detectors miss FR structure
+(NEQ, RAMQ, SIN, FR letterhead, accented ALL-CAPS names).
+
+## Labels (20)
+
+`account_number`, `address`, `card_cvv`, `card_expiry`, `date_of_birth`, `email`, `file_path`, `government_id`,
+`iban`, `ip_address`, `organization`, `password`, `payment_card`, `person`, `phone_number`, `postal_code`,
+`secret`, `sensitive_account_id`, `tax_id`, `username` (41 BIO label ids).
+
+## Intended use
+
+- **Primary:** the always-on detection tier inside the OSSRedact gateway (CPU), and the in-browser redaction
+  workbench (transformers.js).
+- **Also:** on-device / edge PII detection where a ~277 MB INT8 model and CPU latency matter.
+
+> **Use it with a deterministic floor.** As a standalone NER model, recall is below 100%. On this base tier
+> `address` recall is now ~0.93 (no longer weak), but `organization` coverage may still trail the large tier --
+> use the [large tier](https://huggingface.co/ZenSystemAI/ossredact-pii-large) when organization recall matters.
+> OSSRedact's hard guarantee for the catastrophic categories (secrets, cards via Luhn, IBANs, government IDs,
+> emails, IPs, file paths) comes from a Tier-0 floor that runs *independently* of this model.
+
+### Quick start (browser, transformers.js)
+
+```js
+import { pipeline } from '@huggingface/transformers'
+// dtype:'int8' loads onnx/model_int8.onnx -- this repo ships INT8 only (the WASM-native browser format)
+const ner = await pipeline('token-classification', 'ZenSystemAI/ossredact-pii-base', { dtype: 'int8' })
+const out = await ner('Contactez Marie-Eve Tremblay au 514-555-0188; NAS 046 454 286.')
+// out: [{ entity, score, word, start, end }, ...]  -- the document never leaves the page
+```
+
+### Quick start (Python, onnxruntime via optimum)
+
+```python
+from optimum.onnxruntime import ORTModelForTokenClassification
+from transformers import AutoTokenizer, pipeline
+
+tok = AutoTokenizer.from_pretrained("ZenSystemAI/ossredact-pii-base")
+# the repo ships INT8 only (model.int8.onnx at the root); name it explicitly
+model = ORTModelForTokenClassification.from_pretrained("ZenSystemAI/ossredact-pii-base", file_name="model.int8.onnx")
+ner = pipeline("token-classification", model=model, tokenizer=tok)
+print(ner("Reçu de la Caisse Desjardins; IBAN GB82 WEST 1234 5698 7654 32."))
+```
+
+## Training data
+
+A **100% synthetic** French-Québec + English corpus (bank statements, financing forms, email threads, CSV
+exports, `.env` files, code, KYC/tax/SAAQ/RAMQ documents). Every name, SIN, account, card, and secret is
+fabricated. It deliberately includes adversarial cases (ALL-CAPS, NBSP-separated IDs, mixed FR/EN, long unbroken
+lines, look-alike decoys). Same corpus and recipe as the large tier; this is the `v11r9c` revision, trained on the
+cumulative corpus including the organization/address augmentation.
+
+## Evaluation
+
+Synthetic held-out corpus (7,498 rows, 0 train overlap). Privacy metric = full-stack catastrophic DETECTION
+recall; `clean_fp` = over-redaction on no-PII rows.
+
+| tier | catastrophic full-stack DETECTION | all-label recall | precision | clean_fp |
+|------|-----------------------------------|------------------|-----------|----------|
+| **CPU / base (v11r9c)** | **0.9941** | 0.9777 | 0.9139 | 48 / 7498 |
+
+The base model nearly matches the large model on overall recall (trailing by ~1 point: 0.9777 vs 0.9882) at
+~4x lower latency -- the reason it is the always-on tier. `address` recall is now ~0.93 here vs 0.95 on large;
+`organization` coverage may still trail the large tier, so use the
+[large tier](https://huggingface.co/ZenSystemAI/ossredact-pii-large) when organization recall matters.
+
+Training recipe: `xlm-roberta-base` (277 M), batch size 8, learning rate 2e-5, max length 512, 3 epochs,
+`metric_for_best_model=cat_f1`. Figures above are the fp32 `v11r9c` reference; the shipped artifact is the
+per-channel dynamic-INT8 ONNX (~277 MB), pii_argmax 0.967 vs fp32 (parity bar 0.965 -- see the model card's
+INT8 note and validation/RESULT-base-int8-parity-v11r9c.md).
+
+## Limitations
+
+- Trained and validated entirely on **synthetic Québec** data; broader real-world domains are future work.
+- **French and English only** by design.
+- On this base tier, **`organization` coverage may still trail the large tier** (`address` recall is now ~0.93,
+  no longer weak) -- use the large tier when organization recall matters, and in all cases pair with OSSRedact's
+  deterministic floor for the catastrophic categories.
+- Identifier coverage targets **Canadian / Québec** formats. Foreign formats (US ZIP, Brazilian CPF) are not
+  specifically targeted.
+- Recall is below 100%; this model is one layer of a redaction system, not a standalone guarantee.
+
+## License & links
+
+MIT. Part of [OSSRedact](https://github.com/ZenSystemAI/OSSRedact) by ZenSystemAI. The version label `v11rN`
+is the weight revision (an HF revision tag), not part of the repo id.

config.json

@@ -0,0 +1,116 @@
+{
+  "add_cross_attention": false,
+  "architectures": [
+    "XLMRobertaForTokenClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "dtype": "float32",
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": "O",
+    "1": "B-account_number",
+    "2": "I-account_number",
+    "3": "B-address",
+    "4": "I-address",
+    "5": "B-card_cvv",
+    "6": "I-card_cvv",
+    "7": "B-card_expiry",
+    "8": "I-card_expiry",
+    "9": "B-date_of_birth",
+    "10": "I-date_of_birth",
+    "11": "B-email",
+    "12": "I-email",
+    "13": "B-file_path",
+    "14": "I-file_path",
+    "15": "B-government_id",
+    "16": "I-government_id",
+    "17": "B-iban",
+    "18": "I-iban",
+    "19": "B-ip_address",
+    "20": "I-ip_address",
+    "21": "B-organization",
+    "22": "I-organization",
+    "23": "B-password",
+    "24": "I-password",
+    "25": "B-payment_card",
+    "26": "I-payment_card",
+    "27": "B-person",
+    "28": "I-person",
+    "29": "B-phone_number",
+    "30": "I-phone_number",
+    "31": "B-postal_code",
+    "32": "I-postal_code",
+    "33": "B-secret",
+    "34": "I-secret",
+    "35": "B-sensitive_account_id",
+    "36": "I-sensitive_account_id",
+    "37": "B-tax_id",
+    "38": "I-tax_id",
+    "39": "B-username",
+    "40": "I-username"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "is_decoder": false,
+  "label2id": {
+    "B-account_number": 1,
+    "B-address": 3,
+    "B-card_cvv": 5,
+    "B-card_expiry": 7,
+    "B-date_of_birth": 9,
+    "B-email": 11,
+    "B-file_path": 13,
+    "B-government_id": 15,
+    "B-iban": 17,
+    "B-ip_address": 19,
+    "B-organization": 21,
+    "B-password": 23,
+    "B-payment_card": 25,
+    "B-person": 27,
+    "B-phone_number": 29,
+    "B-postal_code": 31,
+    "B-secret": 33,
+    "B-sensitive_account_id": 35,
+    "B-tax_id": 37,
+    "B-username": 39,
+    "I-account_number": 2,
+    "I-address": 4,
+    "I-card_cvv": 6,
+    "I-card_expiry": 8,
+    "I-date_of_birth": 10,
+    "I-email": 12,
+    "I-file_path": 14,
+    "I-government_id": 16,
+    "I-iban": 18,
+    "I-ip_address": 20,
+    "I-organization": 22,
+    "I-password": 24,
+    "I-payment_card": 26,
+    "I-person": 28,
+    "I-phone_number": 30,
+    "I-postal_code": 32,
+    "I-secret": 34,
+    "I-sensitive_account_id": 36,
+    "I-tax_id": 38,
+    "I-username": 40,
+    "O": 0
+  },
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "xlm-roberta",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "output_past": true,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "tie_word_embeddings": true,
+  "transformers_version": "5.12.1",
+  "type_vocab_size": 1,
+  "use_cache": false,
+  "vocab_size": 250002
+}

model.int8.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac9e66bfd3cf9ac50a5d393803bf07ec25970ae7c06b7ce8c2633bea5bc8a462
+size 278409519

onnx/model_int8.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac9e66bfd3cf9ac50a5d393803bf07ec25970ae7c06b7ce8c2633bea5bc8a462
+size 278409519

sentencepiece.bpe.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfc8146abe2a0488e9e2a0c56de7952f7c11ab059eca145a0a727afce0db2865
+size 5069051

special_tokens_map.json

@@ -0,0 +1,15 @@
+{
+  "bos_token": "<s>",
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "mask_token": {
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "unk_token": "<unk>"
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1065c14bf5f3d26a369bb348eae791e137f4b69cc85f846cc14bde0e6c49aca
+size 17098184

tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "add_prefix_space": true,
+  "backend": "tokenizers",
+  "bos_token": "<s>",
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "is_local": false,
+  "local_files_only": false,
+  "mask_token": "<mask>",
+  "model_max_length": 512,
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "tokenizer_class": "XLMRobertaTokenizer",
+  "unk_token": "<unk>"
+}