Publish IrishCore-DiffMask-135M-v1-rc6

.gitattributes

@@ -1,35 +1,5 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
 *.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
 *.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
 *.pt filter=lfs diff=lfs merge=lfs -text
 *.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz 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

.gitignore

@@ -0,0 +1,3 @@
+__pycache__/
+*.pyc
+.venv/

LICENSE

@@ -0,0 +1,73 @@
+Apache License
+Version 2.0, January 2004
+http://www.apache.org/licenses/
+
+TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+1. Definitions.
+
+"License" shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document.
+
+"Licensor" shall mean the copyright owner or entity authorized by the copyright owner that is granting the License.
+
+"Legal Entity" shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, "control" means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity.
+
+"You" (or "Your") shall mean an individual or Legal Entity exercising permissions granted by this License.
+
+"Source" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files.
+
+"Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types.
+
+"Work" shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below).
+
+"Derivative Works" shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof.
+
+"Contribution" shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution."
+
+"Contributor" shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work.
+
+2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
+
+3. Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed.
+
+4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions:
+
+     (a) You must give any other recipients of the Work or Derivative Works a copy of this License; and
+
+     (b) You must cause any modified files to carry prominent notices stating that You changed the files; and
+
+     (c) You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and
+
+     (d) If the Work includes a "NOTICE" text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License.
+
+     You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License.
+
+5. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions.
+
+6. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file.
+
+7. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License.
+
+8. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
+
+9. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
+
+END OF TERMS AND CONDITIONS
+
+APPENDIX: How to apply the Apache License to your work.
+
+To apply the Apache License to your work, attach the following boilerplate notice, with the fields enclosed by brackets "[]" replaced with your own identifying information. (Don't include the brackets!)  The text should be enclosed in the appropriate comment syntax for the file format. We also recommend that a file or class name and description of purpose be included on the same "printed page" as the copyright notice for easier identification within third-party archives.
+
+Copyright [yyyy] [name of copyright owner]
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.

NOTICE

@@ -0,0 +1,13 @@
+IrishCore-DiffMask-135M-v1-rc6
+
+This release is derived from:
+- OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1 (Apache-2.0)
+
+Training and evaluation data used for this derivative included:
+- temsa/OpenMed-Irish-CorePII-TrainMix-v1
+- temsa/OpenMed-Irish-PPSN-Eircode-Spec-v1
+- joelniklaus/mapa (CC-BY-4.0)
+- gretelai/synthetic_pii_finance_multilingual (Apache-2.0)
+
+Additional local synthetic hardening and replay sets were used during model selection.
+Please review upstream licenses and dataset cards before redistributing derivative datasets.

README.md

@@ -0,0 +1,320 @@
+---
+language:
+- en
+- ga
+license: apache-2.0
+library_name: transformers
+pipeline_tag: token-classification
+tags:
+- pii
+- de-identification
+- token-classification
+- ireland
+- irish
+- gaelic
+- diffusion-style
+- denoising
+- ppsn
+- eircode
+- onnx
+- int8
+- dynamic-quantization
+- cpu
+base_model:
+- OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1
+datasets:
+- temsa/OpenMed-Irish-CorePII-TrainMix-v1
+- temsa/OpenMed-Irish-PPSN-Eircode-Spec-v1
+- joelniklaus/mapa
+- gretelai/synthetic_pii_finance_multilingual
+model-index:
+- name: IrishCore-DiffMask-135M-v1-rc6
+  results:
+  - task:
+      type: token-classification
+      name: Irish core PII masking
+    dataset:
+      type: custom
+      name: irish_core_pii_v1
+    metrics:
+    - type: f1
+      name: Overall F1
+      value: 0.9733
+  - task:
+      type: token-classification
+      name: Multilingual PPSN masking
+    dataset:
+      type: custom
+      name: multilingual_ppsn_v1_all
+    metrics:
+    - type: f1
+      name: Overall F1
+      value: 0.9274
+  - task:
+      type: token-classification
+      name: Hardening exact suite
+    dataset:
+      type: custom
+      name: irish_dllm_hardening_exact_v1
+    metrics:
+    - type: f1
+      name: Overall F1
+      value: 0.9756
+---
+
+# IrishCore-DiffMask-135M-v1-rc6
+
+`IrishCore-DiffMask-135M-v1-rc6` is a raw-only Irish PII masking model derived from `OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1`.
+
+It is a small, scanner-free span extractor tuned for:
+
+- `PPSN`
+- `ACCOUNT_NUMBER`
+- `BANK_ROUTING_NUMBER`
+- `CREDIT_DEBIT_CARD`
+- `PASSPORT_NUMBER`
+- `POSTCODE`
+- `PHONE_NUMBER`
+- `EMAIL`
+- `FIRST_NAME`
+- `LAST_NAME`
+- `SWIFT_BIC`
+
+The main target is English plus Irish Gaelic text in citizen-support, public-sector, and HSE-style flows. The repo ships both the full `transformers` checkpoint and a dynamic q8 ONNX artifact for CPU deployment.
+
+## Evaluation Harness
+
+Release gating for this checkpoint uses the deployment-aligned clean single-pass token-span path.
+
+- `experiments/irish_core_span_raw_only/benchmark_multitask.py`
+- `scripts/eval_dllm_release.py --inference-mode clean_single_pass`
+
+These two paths now match exactly on the checked suites.
+The older diffusion-style eval path is retained only as a training diagnostic and is not used for release scoring.
+
+For this release, the best local raw checkpoint is `models/irishcore-diffmask-135m-v1-rc6k-blend-0p30`, and the bundled CPU artifact is `onnx/model_quantized.onnx`.
+
+
+## What "DiffMask" Means Here
+
+This release is not a generative diffusion language model. It is a compact discriminative token-span model trained with a diffusion-style denoising schedule.
+
+The short version:
+
+- **Base OpenMed**: plain BIO token classification
+- **DiffMask**: token-span extraction with token-presence and boundary heads
+- **DiffMask training**: repeated masked denoising over the same sentence
+- **DiffMask inference**: one forward pass, no iterative refinement, no text generation
+
+Concretely:
+
+- The encoder starts from the DistilBERT-family weights inside `OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1`.
+- The model adds three task heads over the encoder hidden states:
+  - a per-label token-presence head
+  - a typed start-boundary head
+  - a typed end-boundary head
+- During training, each input sentence is corrupted multiple times by replacing a random fraction of visible tokens with `[MASK]`.
+- The corruption level follows a short noise schedule from heavy masking to light masking.
+- The same gold spans are learned at every noise level, and the losses are averaged across the denoising passes.
+- At inference time there is no diffusion loop and no rewrite step: the model runs once and a score-only span decoder reconstructs spans from token scores plus typed boundaries.
+
+So the "DLLM" aspect here is the training recipe: repeated masked denoising over text, not autoregressive generation.
+
+## What It Is Not
+
+This model is **not** a full discrete diffusion language model in the LLaDA sense.
+
+A true DLLM would usually have:
+
+- timestep or noise conditioning inside the model
+- iterative denoising at inference time
+- multi-step sequence refinement at runtime
+- text generation or full-sequence reconstruction as a first-class objective
+
+This release does **not** do that.
+
+Instead, it uses the diffusion idea only as a **training-time robustness trick**:
+
+- corrupt the sentence with `[MASK]` at several noise levels
+- train on the same target spans each time
+- average those losses
+
+At runtime, it behaves like a normal fast discriminative extractor.
+
+## Architecture
+
+- Encoder: DistilBERT-size encoder from the OpenMed mLiteClinical 135M base
+- Heads:
+  - token presence per released label
+  - typed start boundary per released label
+  - typed end boundary per released label
+- Decoder:
+  - score-only span decoding from offsets, token continuity, label-specific thresholds, and typed boundaries
+  - no regex candidate extractor
+  - no checksum validator
+  - no scanner layer
+
+The release behavior is fully defined by the weights plus the bundled decoder in `common.py`.
+
+## Training And Inference Flow
+
+Training:
+
+1. tokenize a sentence with gold BIO spans
+2. convert spans into:
+   - token-presence targets
+   - typed start targets
+   - typed end targets
+3. create several noised copies of the same tokenized sentence by masking random visible tokens
+4. run the same encoder+heads on each noised copy
+5. average the losses across those denoising passes
+
+Inference:
+
+1. tokenize the raw text once
+2. run a single forward pass
+3. predict:
+   - which labels are present on each token
+   - where each labeled span starts
+   - where each labeled span ends
+4. decode spans with label-aware thresholds and boundary rules
+5. replace the detected spans with placeholders such as `[PII:PPSN]`
+
+There is no multi-step refinement loop in deployment.
+
+## How It Differs From The Original OpenMed Model
+
+The original `OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1` is a standard `DistilBertForTokenClassification` model:
+
+- one encoder
+- one token-classification head
+- BIO labels such as `B-email`, `I-email`, `B-phone_number`
+- generic token aggregation to recover spans
+
+DiffMask changes two things:
+
+1. **Different supervision**
+   - base OpenMed learns only BIO token labels
+   - DiffMask learns token presence plus typed span boundaries
+
+2. **Different training recipe**
+   - base OpenMed is trained as a standard token classifier
+   - DiffMask is trained on multiple masked-noised views of the same sentence
+
+That makes DiffMask better suited to structured Irish identifiers and mixed PII masking, while still keeping a small encoder and a fast CPU path.
+
+## How It Differs From `rc5` And `rc8`
+
+| Model | Core idea | External scanner/validator | Runtime shape |
+|---|---|---|---|
+| `rc5` | token classifier + repair logic | yes | heavier, decoder-assisted |
+| `rc8` | raw-only token-span model | no | one pass + span decoder |
+| `DiffMask` | raw-only token-span model + denoising training | no | one pass + span decoder |
+
+So DiffMask is closest to `rc8` operationally, but it uses a stronger training recipe.
+
+## Why This Exists
+
+The older `rc5` release still depended on a repair-oriented decoder stack. The public `rc8` release removed that external logic, but it regressed on several structured Irish identifiers. This release keeps the raw-only deployment shape while re-hardening the model on Irish numeric and mixed-PII cases.
+
+## References
+
+Direct implementation references:
+
+- Devlin et al., *BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding*  
+  https://arxiv.org/abs/1810.04805
+- Sanh et al., *DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter*  
+  https://arxiv.org/abs/1910.01108
+- Fu et al., *Boundary Smoothing for Named Entity Recognition*  
+  https://aclanthology.org/2022.acl-long.490/
+- Wang et al., *SPANNER: Named Entity Re-/Recognition as Span Prediction*  
+  https://aclanthology.org/2021.acl-long.558/
+
+Conceptual diffusion-style training references:
+
+- Nie et al., *LLaDA 2.0: Scaling Up Diffusion Language Models to 100B*  
+  https://arxiv.org/abs/2512.15745
+- Gong et al., *Scaling Diffusion Language Models via Adaptation from Autoregressive Models*  
+  https://arxiv.org/abs/2410.17891
+
+These diffusion papers were used as architectural inspiration for the masked noising schedule. This release does **not** implement a generative text diffusion runtime.
+
+## Included Artifacts
+
+- Full `transformers` checkpoint in the repo root
+- Dynamic q8 ONNX export in `onnx/model_quantized.onnx`
+- Unquantized ONNX export in `onnx/model.onnx`
+- `inference_mask.py` for the full checkpoint
+- `inference_mask_onnx.py` for the ONNX q8 path
+- `common.py`, `model.py`, and `multitask_model.py` implementing the release decoder
+- benchmark files in `eval/`
+
+Artifact sizes:
+
+- Full checkpoint: `514 MB` (`model.safetensors`)
+- Dynamic q8 ONNX: `393 MB` (`onnx/model_quantized.onnx`)
+
+## How To Use It
+
+Full checkpoint:
+
+```bash
+uv run python inference_mask.py \
+  --model temsa/IrishCore-DiffMask-135M-v1-rc6 \
+  --min-score 0.5 \
+  --text "My PPSN is 1234567TW, my Eircode is D02 X285, and my phone is 087 123 4567." \
+  --json
+```
+
+Dynamic q8 ONNX:
+
+```bash
+uv run python inference_mask_onnx.py \
+  --model temsa/IrishCore-DiffMask-135M-v1-rc6 \
+  --min-score 0.5 \
+  --text "Please provide your passport NN5123456 and call me on 0851234567." \
+  --json
+```
+
+Both scripts emit explicit placeholders like `[PII:PPSN]` in `masked_text`.
+
+## Q8 Comparison
+
+Deployment-relevant comparison on CPU:
+
+| Model | Core F1 | Edge F1 | Finance F1 | Finance-boundary F1 | User PPSN F1 | GA weak PPSN F1 | Multilingual PPSN F1 | Hardening F1 |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| `rc5` ONNX q8 | 0.9669 | 0.9744 | 0.9362 | 0.8750 | 1.0000 | 1.0000 | 0.9333 | - |
+| `rc8` ONNX q8 | 0.9737 | 1.0000 | 1.0000 | 1.0000 | 1.0000 | 1.0000 | 0.9176 | 0.7059 |
+| `IrishCore-DiffMask-135M-v1-rc6` ONNX q8 | 0.9733 | 0.9500 | 0.9804 | 1.0000 | 0.7500 | 1.0000 | 0.9274 | 0.9756 |
+
+CPU throughput references:
+
+| Suite | `rc5` q8 | `rc8` q8 | `IrishCore-DiffMask-135M-v1-rc6` q8 |
+|---|---:|---:|---:|
+| Irish core short-text path | 33.6193 ex/s | 257.3756 ex/s | 130.3415 ex/s |
+| Multilingual PPSN short-text path | 35.5561 ex/s | 230.5181 ex/s | 200.2041 ex/s |
+| Runtime profile source | 23.8338 ex/s | 179.4708 ex/s | 167.1461 ex/s |
+
+Notes:
+
+- The `rc5` speed references come from its published q8 end-to-end inference stack, which includes its older repair decoder.
+- The `rc8` and `IrishCore-DiffMask-135M-v1-rc6` numbers use the same raw-only token-span ONNX path.
+- A weight-only q4 ONNX experiment was also tried during development, but it was slower than q8 on this CPU and is not shipped.
+
+## Limits
+
+- This is still a compact model. The hardest remaining errors are multilingual PPSN near-miss cases rather than Irish core numeric formats.
+- The release path is intentionally scanner-free. If you need deterministic validation of individual identifier types, add that in your application layer.
+- If you rely on release behavior, use the bundled inference scripts or import `decode_token_presence_segments` from `common.py`.
+
+## License And Attribution
+
+- Release license: Apache-2.0
+- Base model: `OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1`
+- The derivative release remains subject to the attribution terms of the upstream datasets listed above.
+- See `NOTICE`, `training_sources.json`, and `eval/benchmark_summary.json` for provenance and benchmark details.
+
+## Release Note
+
+`rc6` is the first DiffMask package selected after reconciling the evaluation harness. Under the aligned harness, it matches the prior best candidate on fresh holdout, QA feedback exact, UAT exact, multilingual PPSN, edge, finance, and hardening, while improving Irish core on the full checkpoint.

common.py

@@ -0,0 +1,613 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import json
+import tempfile
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+from huggingface_hub import HfApi, hf_hub_download
+from transformers import AutoConfig, AutoTokenizer
+
+TOKENIZER_FILES = [
+    "tokenizer_config.json",
+    "tokenizer.json",
+    "special_tokens_map.json",
+    "vocab.txt",
+    "vocab.json",
+    "merges.txt",
+    "added_tokens.json",
+    "sentencepiece.bpe.model",
+    "spiece.model",
+]
+DEFAULT_LABEL_MAX_SPAN_TOKENS = {
+    # Token-piece limits, not word limits. These need to reflect how the
+    # underlying tokenizer actually fragments compact identifiers.
+    "PPSN": 9,
+    "POSTCODE": 7,
+    "PHONE_NUMBER": 10,
+    "PASSPORT_NUMBER": 8,
+    "BANK_ROUTING_NUMBER": 5,
+    "ACCOUNT_NUMBER": 19,
+    "CREDIT_DEBIT_CARD": 12,
+    "SWIFT_BIC": 8,
+    "EMAIL": 15,
+    "FIRST_NAME": 5,
+    "LAST_NAME": 8,
+}
+DEFAULT_LABEL_MIN_NONSPACE_CHARS = {
+    "PPSN": 8,
+    "POSTCODE": 6,
+    "PHONE_NUMBER": 7,
+    "PASSPORT_NUMBER": 7,
+    "BANK_ROUTING_NUMBER": 6,
+    "ACCOUNT_NUMBER": 6,
+    "CREDIT_DEBIT_CARD": 12,
+    "SWIFT_BIC": 8,
+    "EMAIL": 6,
+    "FIRST_NAME": 2,
+    "LAST_NAME": 2,
+}
+WHITESPACE_BRIDGE_LABELS = {
+    "PPSN",
+    "POSTCODE",
+    "PHONE_NUMBER",
+    "PASSPORT_NUMBER",
+    "BANK_ROUTING_NUMBER",
+    "ACCOUNT_NUMBER",
+    "CREDIT_DEBIT_CARD",
+    "SWIFT_BIC",
+}
+SIMPLE_PUNCT_BRIDGE_LABELS = {
+    "PHONE_NUMBER",
+    "BANK_ROUTING_NUMBER",
+    "ACCOUNT_NUMBER",
+    "CREDIT_DEBIT_CARD",
+}
+MIN_CHAR_FALLBACK_LABELS = {
+    "PHONE_NUMBER",
+    "BANK_ROUTING_NUMBER",
+    "ACCOUNT_NUMBER",
+    "CREDIT_DEBIT_CARD",
+    "EMAIL",
+}
+CONSERVATIVE_BOUNDARY_REFINEMENT_LABELS = {
+    "PPSN",
+    "POSTCODE",
+    "PHONE_NUMBER",
+    "PASSPORT_NUMBER",
+    "BANK_ROUTING_NUMBER",
+    "ACCOUNT_NUMBER",
+    "CREDIT_DEBIT_CARD",
+    "SWIFT_BIC",
+    "EMAIL",
+}
+OUTPUT_PRIORITY = {
+    "PPSN": 0,
+    "PASSPORT_NUMBER": 1,
+    "ACCOUNT_NUMBER": 2,
+    "BANK_ROUTING_NUMBER": 3,
+    "CREDIT_DEBIT_CARD": 4,
+    "PHONE_NUMBER": 5,
+    "SWIFT_BIC": 6,
+    "POSTCODE": 7,
+    "EMAIL": 8,
+    "FIRST_NAME": 9,
+    "LAST_NAME": 10,
+}
+
+
+def normalize_entity_name(label: str) -> str:
+    label = (label or "").strip()
+    if label.startswith("B-") or label.startswith("I-"):
+        label = label[2:]
+    return label.upper()
+
+
+def _sanitize_tokenizer_dir(tokenizer_path: Path) -> str:
+    tokenizer_cfg_path = tokenizer_path / "tokenizer_config.json"
+    if not tokenizer_cfg_path.exists():
+        return str(tokenizer_path)
+    data = json.loads(tokenizer_cfg_path.read_text(encoding="utf-8"))
+    if "fix_mistral_regex" not in data:
+        return str(tokenizer_path)
+    tmpdir = Path(tempfile.mkdtemp(prefix="openmed_span_tokenizer_"))
+    keep = set(TOKENIZER_FILES)
+    for child in tokenizer_path.iterdir():
+        if child.is_file() and child.name in keep:
+            (tmpdir / child.name).write_bytes(child.read_bytes())
+    data.pop("fix_mistral_regex", None)
+    (tmpdir / "tokenizer_config.json").write_text(json.dumps(data, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+    return str(tmpdir)
+
+
+def safe_auto_tokenizer(tokenizer_ref: str):
+    tokenizer_path = Path(tokenizer_ref)
+    if tokenizer_path.exists():
+        tokenizer_ref = _sanitize_tokenizer_dir(tokenizer_path)
+    else:
+        api = HfApi()
+        files = set(api.list_repo_files(repo_id=tokenizer_ref, repo_type="model"))
+        tmpdir = Path(tempfile.mkdtemp(prefix="openmed_remote_span_tokenizer_"))
+        copied = False
+        for name in TOKENIZER_FILES:
+            if name not in files:
+                continue
+            src = hf_hub_download(repo_id=tokenizer_ref, filename=name, repo_type="model")
+            (tmpdir / Path(name).name).write_bytes(Path(src).read_bytes())
+            copied = True
+        if copied:
+            tokenizer_ref = _sanitize_tokenizer_dir(tmpdir)
+
+    try:
+        return AutoTokenizer.from_pretrained(tokenizer_ref, use_fast=True, fix_mistral_regex=True)
+    except Exception:
+        pass
+    try:
+        return AutoTokenizer.from_pretrained(tokenizer_ref, use_fast=True, fix_mistral_regex=False)
+    except TypeError:
+        pass
+    try:
+        return AutoTokenizer.from_pretrained(tokenizer_ref, use_fast=True)
+    except Exception:
+        return AutoTokenizer.from_pretrained(tokenizer_ref, use_fast=False)
+
+
+def label_names_from_config(config) -> list[str]:
+    names = list(getattr(config, "span_label_names", []))
+    if not names:
+        raise ValueError("Missing span_label_names in config")
+    return [normalize_entity_name(name) for name in names]
+
+
+def label_thresholds_from_config(config, default_threshold: float) -> dict[str, float]:
+    raw = getattr(config, "span_label_thresholds", None) or {}
+    out = {normalize_entity_name(key): float(value) for key, value in raw.items()}
+    for label in label_names_from_config(config):
+        out.setdefault(label, float(default_threshold))
+    return out
+
+
+def token_label_thresholds_from_config(config, default_threshold: float) -> dict[str, float]:
+    raw = getattr(config, "token_label_thresholds", None) or {}
+    out = {normalize_entity_name(key): float(value) for key, value in raw.items()}
+    for label in label_names_from_config(config):
+        out.setdefault(label, float(default_threshold))
+    return out
+
+
+def token_extend_thresholds_from_config(config, default_fraction: float = 0.6) -> dict[str, float]:
+    raw = getattr(config, "token_extend_thresholds", None) or {}
+    out = {normalize_entity_name(key): float(value) for key, value in raw.items()}
+    for label in label_names_from_config(config):
+        out.setdefault(label, max(0.0, min(1.0, float(token_label_thresholds_from_config(config, 0.5).get(label, 0.5)) * default_fraction)))
+    return out
+
+
+def boundary_label_thresholds_from_config(config, default_threshold: float = 0.0) -> dict[str, float]:
+    raw = getattr(config, "boundary_label_thresholds", None) or {}
+    out = {normalize_entity_name(key): float(value) for key, value in raw.items()}
+    for label in label_names_from_config(config):
+        out.setdefault(label, float(default_threshold))
+    return out
+
+
+def label_max_span_tokens_from_config(config) -> dict[str, int]:
+    raw = getattr(config, "span_label_max_span_tokens", None) or {}
+    out = {normalize_entity_name(key): int(value) for key, value in raw.items()}
+    for label, value in DEFAULT_LABEL_MAX_SPAN_TOKENS.items():
+        out.setdefault(label, value)
+    for label in label_names_from_config(config):
+        out.setdefault(label, 8)
+    return out
+
+
+def label_min_nonspace_chars_from_config(config) -> dict[str, int]:
+    raw = getattr(config, "span_label_min_nonspace_chars", None) or {}
+    out = {normalize_entity_name(key): int(value) for key, value in raw.items()}
+    for label, value in DEFAULT_LABEL_MIN_NONSPACE_CHARS.items():
+        out.setdefault(label, value)
+    for label in label_names_from_config(config):
+        out.setdefault(label, 1)
+    return out
+
+
+def overlaps(a: dict, b: dict) -> bool:
+    return not (a["end"] <= b["start"] or b["end"] <= a["start"])
+
+
+def dedupe_spans(spans: list[dict]) -> list[dict]:
+    ordered = sorted(
+        spans,
+        key=lambda item: (-float(item.get("score", 0.0)), item["start"], item["end"], OUTPUT_PRIORITY.get(item["label"], 99)),
+    )
+    kept = []
+    for span in ordered:
+        if any(overlaps(span, other) for other in kept):
+            continue
+        kept.append(span)
+    kept.sort(key=lambda item: (item["start"], item["end"], OUTPUT_PRIORITY.get(item["label"], 99)))
+    return kept
+
+
+def _valid_offset(offset: tuple[int, int]) -> bool:
+    return bool(offset) and offset[1] > offset[0]
+
+
+def _has_skippable_bridge(text: str, left: tuple[int, int], right: tuple[int, int], label: str) -> bool:
+    bridge = text[int(left[1]) : int(right[0])]
+    if bridge == "":
+        return True
+    if label == "PPSN" and bridge.isspace():
+        next_token = _token_text(text, right).strip()
+        return 0 < len(next_token) <= 2 and next_token.isalnum()
+    if label in WHITESPACE_BRIDGE_LABELS and bridge.isspace():
+        return True
+    if label in SIMPLE_PUNCT_BRIDGE_LABELS:
+        normalized = bridge.replace("\u00A0", " ").replace("\u202F", " ").strip()
+        if normalized == "-":
+            return True
+    return False
+
+
+def _has_left_extension_bridge(text: str, left: tuple[int, int], right: tuple[int, int]) -> bool:
+    bridge = text[int(left[1]) : int(right[0])]
+    return bridge == ""
+
+
+def _nonspace_length(text: str, start: int, end: int) -> int:
+    return sum(0 if ch.isspace() else 1 for ch in text[int(start) : int(end)])
+
+
+def _is_simple_punct_token(text: str, offset: tuple[int, int], label: str) -> bool:
+    if label not in SIMPLE_PUNCT_BRIDGE_LABELS or not _valid_offset(offset):
+        return False
+    token_text = text[int(offset[0]) : int(offset[1])].replace("\u00A0", " ").replace("\u202F", " ").strip()
+    return token_text == "-"
+
+
+def _token_text(text: str, offset: tuple[int, int]) -> str:
+    return text[int(offset[0]) : int(offset[1])]
+
+
+def _is_short_alnum_token(text: str, offset: tuple[int, int], max_len: int = 4) -> bool:
+    token_text = _token_text(text, offset).strip()
+    return 0 < len(token_text) <= max_len and token_text.isalnum()
+
+
+def _rescue_structured_start(
+    text: str,
+    offsets: list[tuple[int, int]],
+    valid: list[bool],
+    token_scores: np.ndarray,
+    start_scores: np.ndarray,
+    label: str,
+    label_index: int,
+    threshold: float,
+    boundary_threshold: float,
+    start_idx: int,
+    end_idx: int,
+) -> int | None:
+    if label not in {"ACCOUNT_NUMBER", "CREDIT_DEBIT_CARD"}:
+        return None
+    segment_text = text[int(offsets[start_idx][0]) : int(offsets[end_idx][1])]
+    if label == "ACCOUNT_NUMBER" and not any(ch.isspace() for ch in segment_text):
+        return None
+    best_idx = None
+    best_score = -1.0
+    for cand_idx in range(start_idx, end_idx + 1):
+        if not valid[cand_idx]:
+            continue
+        token_score = float(token_scores[cand_idx, label_index])
+        start_score = float(start_scores[cand_idx, label_index])
+        if token_score < threshold or start_score < boundary_threshold:
+            continue
+        token_text = _token_text(text, offsets[cand_idx]).strip()
+        score = start_score + 0.2 * token_score
+        if label == "ACCOUNT_NUMBER":
+            next_text = _token_text(text, offsets[cand_idx + 1]).strip() if cand_idx + 1 <= end_idx and valid[cand_idx + 1] else ""
+            if token_text.upper() == "I" and next_text.upper() == "E":
+                score += 1.0
+            elif token_text.upper().startswith("IE"):
+                score += 0.6
+        elif label == "CREDIT_DEBIT_CARD" and token_text.isdigit():
+            score += 0.3
+        if score > best_score:
+            best_idx = cand_idx
+            best_score = score
+    return best_idx
+
+
+def _rescue_email_outer_span(span_text: str, outer_text: str) -> bool:
+    if "@" not in span_text or " " in outer_text:
+        return False
+    if "@" not in outer_text:
+        return False
+    _, _, span_domain = span_text.partition("@")
+    _, _, outer_domain = outer_text.partition("@")
+    if "." in span_domain and not span_text.endswith("@"):
+        return False
+    return "." in outer_domain and not outer_text.endswith("@")
+
+
+def _rescue_iban_tail(text: str, offsets: list[tuple[int, int]], valid: list[bool], start_idx: int, end_idx: int) -> int:
+    next_idx = end_idx + 1
+    span_text = text[int(offsets[start_idx][0]) : int(offsets[end_idx][1])]
+    if not any(ch.isspace() for ch in span_text):
+        return end_idx
+    compact = "".join(ch for ch in span_text if not ch.isspace())
+    if not compact.upper().startswith("IE"):
+        return end_idx
+    while next_idx < len(offsets) and valid[next_idx]:
+        if not _has_skippable_bridge(text, offsets[end_idx], offsets[next_idx], "ACCOUNT_NUMBER"):
+            break
+        if not _is_short_alnum_token(text, offsets[next_idx]):
+            break
+        end_idx = next_idx
+        span_text = text[int(offsets[start_idx][0]) : int(offsets[end_idx][1])]
+        compact = "".join(ch for ch in span_text if not ch.isspace())
+        if len(compact) >= 22:
+            break
+        next_idx += 1
+    return end_idx
+
+
+def decode_span_logits(
+    text: str,
+    offsets: list[tuple[int, int]],
+    start_scores: np.ndarray,
+    end_scores: np.ndarray,
+    label_names: list[str],
+    default_threshold: float,
+    label_thresholds: dict[str, float] | None = None,
+    label_max_span_tokens: dict[str, int] | None = None,
+) -> list[dict]:
+    thresholds = {label: float(default_threshold) for label in label_names}
+    if label_thresholds:
+        thresholds.update({normalize_entity_name(key): float(value) for key, value in label_thresholds.items()})
+    max_tokens = dict(DEFAULT_LABEL_MAX_SPAN_TOKENS)
+    if label_max_span_tokens:
+        max_tokens.update({normalize_entity_name(key): int(value) for key, value in label_max_span_tokens.items()})
+
+    spans: list[dict] = []
+    for label_index, label in enumerate(label_names):
+        threshold = thresholds.get(label, float(default_threshold))
+        max_span = max_tokens.get(label, 8)
+        start_candidates = [idx for idx in range(len(offsets)) if _valid_offset(offsets[idx]) and float(start_scores[idx, label_index]) >= threshold]
+        for start_idx in start_candidates:
+            best = None
+            for end_idx in range(start_idx, min(len(offsets), start_idx + max_span)):
+                if not _valid_offset(offsets[end_idx]):
+                    continue
+                end_score = float(end_scores[end_idx, label_index])
+                if end_score < threshold:
+                    continue
+                score = min(float(start_scores[start_idx, label_index]), end_score)
+                if best is None or score > best["score"]:
+                    best = {
+                        "label": label,
+                        "start": int(offsets[start_idx][0]),
+                        "end": int(offsets[end_idx][1]),
+                        "score": score,
+                    }
+            if best is not None and best["start"] < best["end"]:
+                best["text"] = text[best["start"]:best["end"]]
+                spans.append(best)
+    return dedupe_spans(spans)
+
+
+def decode_token_presence_segments(
+    text: str,
+    offsets: list[tuple[int, int]],
+    token_scores: np.ndarray,
+    label_names: list[str],
+    default_threshold: float,
+    label_thresholds: dict[str, float] | None = None,
+    label_extend_thresholds: dict[str, float] | None = None,
+    label_max_span_tokens: dict[str, int] | None = None,
+    label_min_nonspace_chars: dict[str, int] | None = None,
+    boundary_label_thresholds: dict[str, float] | None = None,
+    start_scores: np.ndarray | None = None,
+    end_scores: np.ndarray | None = None,
+) -> list[dict]:
+    thresholds = {label: float(default_threshold) for label in label_names}
+    if label_thresholds:
+        thresholds.update({normalize_entity_name(key): float(value) for key, value in label_thresholds.items()})
+    extend_thresholds = {label: max(0.0, min(1.0, thresholds[label] * 0.6)) for label in label_names}
+    if label_extend_thresholds:
+        extend_thresholds.update({normalize_entity_name(key): float(value) for key, value in label_extend_thresholds.items()})
+    max_tokens = dict(DEFAULT_LABEL_MAX_SPAN_TOKENS)
+    if label_max_span_tokens:
+        max_tokens.update({normalize_entity_name(key): int(value) for key, value in label_max_span_tokens.items()})
+    min_nonspace_chars = dict(DEFAULT_LABEL_MIN_NONSPACE_CHARS)
+    if label_min_nonspace_chars:
+        min_nonspace_chars.update({normalize_entity_name(key): int(value) for key, value in label_min_nonspace_chars.items()})
+    boundary_thresholds = {label: 0.0 for label in label_names}
+    if boundary_label_thresholds:
+        boundary_thresholds.update({normalize_entity_name(key): float(value) for key, value in boundary_label_thresholds.items()})
+
+    spans: list[dict] = []
+    valid = [_valid_offset(offset) for offset in offsets]
+    num_tokens = len(offsets)
+    for label_index, label in enumerate(label_names):
+        threshold = thresholds.get(label, float(default_threshold))
+        extend_threshold = min(threshold, extend_thresholds.get(label, threshold))
+        max_span = max_tokens.get(label, 8)
+        idx = 0
+        while idx < num_tokens:
+            if not valid[idx] or float(token_scores[idx, label_index]) < threshold:
+                idx += 1
+                continue
+            start_idx = idx
+            end_idx = idx
+            outer_start_idx = start_idx
+            outer_end_idx = end_idx
+            while end_idx + 1 < num_tokens and valid[end_idx + 1] and float(token_scores[end_idx + 1, label_index]) >= threshold and (end_idx + 1 - start_idx + 1) <= max_span:
+                end_idx += 1
+            while (
+                start_idx - 1 >= 0
+                and valid[start_idx - 1]
+                and _has_left_extension_bridge(text, offsets[start_idx - 1], offsets[start_idx])
+                and float(token_scores[start_idx - 1, label_index]) >= extend_threshold
+                and (end_idx - (start_idx - 1) + 1) <= max_span
+            ):
+                start_idx -= 1
+            while end_idx + 1 < num_tokens:
+                next_idx = end_idx + 1
+                if not valid[next_idx]:
+                    break
+                if (
+                    _has_skippable_bridge(text, offsets[end_idx], offsets[next_idx], label)
+                    and float(token_scores[next_idx, label_index]) >= extend_threshold
+                    and (next_idx - start_idx + 1) <= max_span
+                ):
+                    end_idx = next_idx
+                    continue
+                if (
+                    _is_simple_punct_token(text, offsets[next_idx], label)
+                    and next_idx + 1 < num_tokens
+                    and valid[next_idx + 1]
+                    and _has_skippable_bridge(text, offsets[end_idx], offsets[next_idx], label)
+                    and _has_skippable_bridge(text, offsets[next_idx], offsets[next_idx + 1], label)
+                    and float(token_scores[next_idx + 1, label_index]) >= extend_threshold
+                    and ((next_idx + 1) - start_idx + 1) <= max_span
+                ):
+                    end_idx = next_idx + 1
+                    continue
+                break
+            outer_start_idx = start_idx
+            outer_end_idx = end_idx
+            presence_slice = token_scores[start_idx : end_idx + 1, label_index]
+            score = float(presence_slice.mean())
+            out_start_idx = start_idx
+            out_end_idx = end_idx
+            if start_scores is not None and end_scores is not None:
+                refine_window = min(3, end_idx - start_idx + 1)
+                start_window = start_scores[start_idx : start_idx + refine_window, label_index]
+                best_start_rel = int(np.argmax(start_window))
+                best_start_idx = start_idx + best_start_rel
+                end_window_start = max(best_start_idx, end_idx - refine_window + 1)
+                end_window = end_scores[end_window_start : end_idx + 1, label_index]
+                best_end_rel = int(np.argmax(end_window))
+                best_end_idx = end_window_start + best_end_rel
+                if (
+                    float(start_scores[best_start_idx, label_index]) < boundary_thresholds.get(label, 0.0)
+                    or float(end_scores[best_end_idx, label_index]) < boundary_thresholds.get(label, 0.0)
+                ):
+                    rescued_start_idx = _rescue_structured_start(
+                        text,
+                        offsets,
+                        valid,
+                        token_scores,
+                        start_scores,
+                        label,
+                        label_index,
+                        threshold,
+                        boundary_thresholds.get(label, 0.0),
+                        start_idx,
+                        end_idx,
+                    )
+                    if rescued_start_idx is not None:
+                        out_start_idx = rescued_start_idx
+                        out_end_idx = end_idx
+                    else:
+                        idx = end_idx + 1
+                        continue
+                else:
+                    out_start_idx = best_start_idx
+                    out_end_idx = best_end_idx
+                    if label in CONSERVATIVE_BOUNDARY_REFINEMENT_LABELS and (
+                        best_start_idx != start_idx or best_end_idx != end_idx
+                    ):
+                        outer_boundary = min(float(start_scores[start_idx, label_index]), float(end_scores[end_idx, label_index]))
+                        refined_boundary = min(
+                            float(start_scores[best_start_idx, label_index]),
+                            float(end_scores[best_end_idx, label_index]),
+                        )
+                        if refined_boundary < outer_boundary + 0.08:
+                            out_start_idx = start_idx
+                            out_end_idx = end_idx
+                score = (
+                    0.65 * score
+                    + 0.175 * float(start_scores[out_start_idx, label_index])
+                    + 0.175 * float(end_scores[out_end_idx, label_index])
+                )
+            min_chars = int(min_nonspace_chars.get(label, 1))
+            if _nonspace_length(text, offsets[out_start_idx][0], offsets[out_end_idx][1]) < min_chars:
+                if (
+                    label in MIN_CHAR_FALLBACK_LABELS
+                    and (out_start_idx != start_idx or out_end_idx != end_idx)
+                    and _nonspace_length(text, offsets[start_idx][0], offsets[end_idx][1]) >= min_chars
+                ):
+                    out_start_idx = start_idx
+                    out_end_idx = end_idx
+                else:
+                    idx = end_idx + 1
+                    continue
+            if label == "ACCOUNT_NUMBER":
+                out_end_idx = _rescue_iban_tail(text, offsets, valid, out_start_idx, out_end_idx)
+            span_text = text[int(offsets[out_start_idx][0]) : int(offsets[out_end_idx][1])]
+            outer_text = text[int(offsets[outer_start_idx][0]) : int(offsets[outer_end_idx][1])]
+            if label == "EMAIL" and _rescue_email_outer_span(span_text, outer_text):
+                out_start_idx = outer_start_idx
+                out_end_idx = outer_end_idx
+                span_text = outer_text
+            if label in {"FIRST_NAME", "LAST_NAME"} and any(ch.isdigit() for ch in span_text):
+                idx = end_idx + 1
+                continue
+            spans.append(
+                {
+                    "label": label,
+                    "start": int(offsets[out_start_idx][0]),
+                    "end": int(offsets[out_end_idx][1]),
+                    "score": score,
+                    "text": span_text,
+                }
+            )
+            idx = end_idx + 1
+    return dedupe_spans(spans)
+
+
+def load_onnx_session(model_ref: str, onnx_file: str = "model_quantized.onnx", onnx_subfolder: str = "onnx"):
+    import onnxruntime as ort
+
+    model_path = Path(model_ref)
+    if model_path.exists():
+        candidates = []
+        if onnx_subfolder:
+            candidates.append(model_path / onnx_subfolder / onnx_file)
+        candidates.append(model_path / onnx_file)
+        onnx_path = next((path for path in candidates if path.exists()), candidates[0])
+        config = AutoConfig.from_pretrained(model_ref)
+        tokenizer = safe_auto_tokenizer(model_ref)
+    else:
+        remote_name = f"{onnx_subfolder}/{onnx_file}" if onnx_subfolder else onnx_file
+        onnx_path = Path(hf_hub_download(repo_id=model_ref, filename=remote_name, repo_type="model"))
+        config = AutoConfig.from_pretrained(model_ref)
+        tokenizer = safe_auto_tokenizer(model_ref)
+    session = ort.InferenceSession(str(onnx_path), providers=["CPUExecutionProvider"])
+    return session, tokenizer, config
+
+
+def run_onnx(session, encoded: dict[str, Any]) -> tuple[np.ndarray, np.ndarray]:
+    feed = {}
+    input_names = {item.name for item in session.get_inputs()}
+    for key, value in encoded.items():
+        if key == "offset_mapping":
+            continue
+        if key in input_names:
+            feed[key] = value
+    outputs = session.run(None, feed)
+    return outputs[0], outputs[1]
+
+
+def run_onnx_all(session, encoded: dict[str, Any]) -> list[np.ndarray]:
+    feed = {}
+    input_names = {item.name for item in session.get_inputs()}
+    for key, value in encoded.items():
+        if key == "offset_mapping":
+            continue
+        if key in input_names:
+            feed[key] = value
+    return session.run(None, feed)

config.json

@@ -0,0 +1,338 @@
+{
+  "activation": "gelu",
+  "architectures": [
+    "IrishCoreTokenSpanModel"
+  ],
+  "attention_dropout": 0.1,
+  "boundary_focal_gamma": 2.0,
+  "boundary_hard_fraction": 0.25,
+  "boundary_label_thresholds": {
+    "ACCOUNT_NUMBER": 0.1,
+    "BANK_ROUTING_NUMBER": 0.25,
+    "CREDIT_DEBIT_CARD": 0.25,
+    "EMAIL": 0.2,
+    "FIRST_NAME": 0.1,
+    "LAST_NAME": 0.1,
+    "PASSPORT_NUMBER": 0.25,
+    "PHONE_NUMBER": 0.25,
+    "POSTCODE": 0.4,
+    "PPSN": 0.35,
+    "SWIFT_BIC": 0.25
+  },
+  "boundary_loss_weight": 1.0,
+  "decoder_profile": "rc3_focusv3_email_extend015_passport075",
+  "dim": 768,
+  "dropout": 0.1,
+  "dtype": "float32",
+  "hidden_dim": 3072,
+  "id2label": {
+    "0": "O",
+    "1": "B-account_number",
+    "2": "B-age",
+    "3": "B-api_key",
+    "4": "B-bank_routing_number",
+    "5": "B-biometric_identifier",
+    "6": "B-blood_type",
+    "7": "B-certificate_license_number",
+    "8": "B-city",
+    "9": "B-company_name",
+    "10": "B-coordinate",
+    "11": "B-country",
+    "12": "B-county",
+    "13": "B-credit_debit_card",
+    "14": "B-customer_id",
+    "15": "B-cvv",
+    "16": "B-date",
+    "17": "B-date_of_birth",
+    "18": "B-date_time",
+    "19": "B-device_identifier",
+    "20": "B-education_level",
+    "21": "B-email",
+    "22": "B-employee_id",
+    "23": "B-employment_status",
+    "24": "B-fax_number",
+    "25": "B-first_name",
+    "26": "B-gender",
+    "27": "B-health_plan_beneficiary_number",
+    "28": "B-http_cookie",
+    "29": "B-ipv4",
+    "30": "B-ipv6",
+    "31": "B-language",
+    "32": "B-last_name",
+    "33": "B-license_plate",
+    "34": "B-mac_address",
+    "35": "B-medical_record_number",
+    "36": "B-occupation",
+    "37": "B-password",
+    "38": "B-phone_number",
+    "39": "B-pin",
+    "40": "B-political_view",
+    "41": "B-postcode",
+    "42": "B-race_ethnicity",
+    "43": "B-religious_belief",
+    "44": "B-sexuality",
+    "45": "B-ssn",
+    "46": "B-state",
+    "47": "B-street_address",
+    "48": "B-swift_bic",
+    "49": "B-tax_id",
+    "50": "B-time",
+    "51": "B-unique_id",
+    "52": "B-url",
+    "53": "B-user_name",
+    "54": "B-vehicle_identifier",
+    "55": "I-account_number",
+    "56": "I-api_key",
+    "57": "I-biometric_identifier",
+    "58": "I-blood_type",
+    "59": "I-certificate_license_number",
+    "60": "I-city",
+    "61": "I-company_name",
+    "62": "I-coordinate",
+    "63": "I-country",
+    "64": "I-county",
+    "65": "I-credit_debit_card",
+    "66": "I-customer_id",
+    "67": "I-date",
+    "68": "I-date_of_birth",
+    "69": "I-date_time",
+    "70": "I-device_identifier",
+    "71": "I-education_level",
+    "72": "I-email",
+    "73": "I-employee_id",
+    "74": "I-employment_status",
+    "75": "I-fax_number",
+    "76": "I-first_name",
+    "77": "I-gender",
+    "78": "I-health_plan_beneficiary_number",
+    "79": "I-http_cookie",
+    "80": "I-ipv4",
+    "81": "I-ipv6",
+    "82": "I-language",
+    "83": "I-last_name",
+    "84": "I-license_plate",
+    "85": "I-mac_address",
+    "86": "I-medical_record_number",
+    "87": "I-occupation",
+    "88": "I-password",
+    "89": "I-phone_number",
+    "90": "I-pin",
+    "91": "I-political_view",
+    "92": "I-postcode",
+    "93": "I-race_ethnicity",
+    "94": "I-religious_belief",
+    "95": "I-sexuality",
+    "96": "I-ssn",
+    "97": "I-state",
+    "98": "I-street_address",
+    "99": "I-swift_bic",
+    "100": "I-tax_id",
+    "101": "I-time",
+    "102": "I-unique_id",
+    "103": "I-url",
+    "104": "I-user_name",
+    "105": "I-vehicle_identifier",
+    "106": "B-PPSN",
+    "107": "I-PPSN",
+    "108": "B-PASSPORT_NUMBER",
+    "109": "I-PASSPORT_NUMBER",
+    "110": "I-bank_routing_number"
+  },
+  "initializer_range": 0.02,
+  "label2id": {
+    "B-PASSPORT_NUMBER": 108,
+    "B-PPSN": 106,
+    "B-account_number": 1,
+    "B-age": 2,
+    "B-api_key": 3,
+    "B-bank_routing_number": 4,
+    "B-biometric_identifier": 5,
+    "B-blood_type": 6,
+    "B-certificate_license_number": 7,
+    "B-city": 8,
+    "B-company_name": 9,
+    "B-coordinate": 10,
+    "B-country": 11,
+    "B-county": 12,
+    "B-credit_debit_card": 13,
+    "B-customer_id": 14,
+    "B-cvv": 15,
+    "B-date": 16,
+    "B-date_of_birth": 17,
+    "B-date_time": 18,
+    "B-device_identifier": 19,
+    "B-education_level": 20,
+    "B-email": 21,
+    "B-employee_id": 22,
+    "B-employment_status": 23,
+    "B-fax_number": 24,
+    "B-first_name": 25,
+    "B-gender": 26,
+    "B-health_plan_beneficiary_number": 27,
+    "B-http_cookie": 28,
+    "B-ipv4": 29,
+    "B-ipv6": 30,
+    "B-language": 31,
+    "B-last_name": 32,
+    "B-license_plate": 33,
+    "B-mac_address": 34,
+    "B-medical_record_number": 35,
+    "B-occupation": 36,
+    "B-password": 37,
+    "B-phone_number": 38,
+    "B-pin": 39,
+    "B-political_view": 40,
+    "B-postcode": 41,
+    "B-race_ethnicity": 42,
+    "B-religious_belief": 43,
+    "B-sexuality": 44,
+    "B-ssn": 45,
+    "B-state": 46,
+    "B-street_address": 47,
+    "B-swift_bic": 48,
+    "B-tax_id": 49,
+    "B-time": 50,
+    "B-unique_id": 51,
+    "B-url": 52,
+    "B-user_name": 53,
+    "B-vehicle_identifier": 54,
+    "I-PASSPORT_NUMBER": 109,
+    "I-PPSN": 107,
+    "I-account_number": 55,
+    "I-api_key": 56,
+    "I-bank_routing_number": 110,
+    "I-biometric_identifier": 57,
+    "I-blood_type": 58,
+    "I-certificate_license_number": 59,
+    "I-city": 60,
+    "I-company_name": 61,
+    "I-coordinate": 62,
+    "I-country": 63,
+    "I-county": 64,
+    "I-credit_debit_card": 65,
+    "I-customer_id": 66,
+    "I-date": 67,
+    "I-date_of_birth": 68,
+    "I-date_time": 69,
+    "I-device_identifier": 70,
+    "I-education_level": 71,
+    "I-email": 72,
+    "I-employee_id": 73,
+    "I-employment_status": 74,
+    "I-fax_number": 75,
+    "I-first_name": 76,
+    "I-gender": 77,
+    "I-health_plan_beneficiary_number": 78,
+    "I-http_cookie": 79,
+    "I-ipv4": 80,
+    "I-ipv6": 81,
+    "I-language": 82,
+    "I-last_name": 83,
+    "I-license_plate": 84,
+    "I-mac_address": 85,
+    "I-medical_record_number": 86,
+    "I-occupation": 87,
+    "I-password": 88,
+    "I-phone_number": 89,
+    "I-pin": 90,
+    "I-political_view": 91,
+    "I-postcode": 92,
+    "I-race_ethnicity": 93,
+    "I-religious_belief": 94,
+    "I-sexuality": 95,
+    "I-ssn": 96,
+    "I-state": 97,
+    "I-street_address": 98,
+    "I-swift_bic": 99,
+    "I-tax_id": 100,
+    "I-time": 101,
+    "I-unique_id": 102,
+    "I-url": 103,
+    "I-user_name": 104,
+    "I-vehicle_identifier": 105,
+    "O": 0
+  },
+  "max_position_embeddings": 512,
+  "model_type": "distilbert",
+  "n_heads": 12,
+  "n_layers": 6,
+  "num_span_labels": 11,
+  "output_past": true,
+  "pad_token_id": 0,
+  "qa_dropout": 0.1,
+  "seq_classif_dropout": 0.2,
+  "sinusoidal_pos_embds": false,
+  "span_label_max_span_tokens": {
+    "ACCOUNT_NUMBER": 19,
+    "BANK_ROUTING_NUMBER": 6,
+    "CREDIT_DEBIT_CARD": 13,
+    "EMAIL": 16,
+    "FIRST_NAME": 5,
+    "LAST_NAME": 8,
+    "PASSPORT_NUMBER": 9,
+    "PHONE_NUMBER": 10,
+    "POSTCODE": 8,
+    "PPSN": 9,
+    "SWIFT_BIC": 8
+  },
+  "span_label_names": [
+    "ACCOUNT_NUMBER",
+    "BANK_ROUTING_NUMBER",
+    "CREDIT_DEBIT_CARD",
+    "EMAIL",
+    "FIRST_NAME",
+    "LAST_NAME",
+    "PASSPORT_NUMBER",
+    "PHONE_NUMBER",
+    "POSTCODE",
+    "PPSN",
+    "SWIFT_BIC"
+  ],
+  "span_label_thresholds": {
+    "ACCOUNT_NUMBER": 0.5,
+    "BANK_ROUTING_NUMBER": 0.5,
+    "CREDIT_DEBIT_CARD": 0.5,
+    "EMAIL": 0.5,
+    "FIRST_NAME": 0.5,
+    "LAST_NAME": 0.5,
+    "PASSPORT_NUMBER": 0.5,
+    "PHONE_NUMBER": 0.5,
+    "POSTCODE": 0.5,
+    "PPSN": 0.5,
+    "SWIFT_BIC": 0.5
+  },
+  "span_positive_weight": 6.0,
+  "tie_weights_": true,
+  "token_extend_thresholds": {
+    "ACCOUNT_NUMBER": 0.08,
+    "BANK_ROUTING_NUMBER": 0.3,
+    "CREDIT_DEBIT_CARD": 0.3,
+    "EMAIL": 0.15,
+    "FIRST_NAME": 0.3,
+    "LAST_NAME": 0.3,
+    "PASSPORT_NUMBER": 0.15,
+    "PHONE_NUMBER": 0.15,
+    "POSTCODE": 0.3,
+    "PPSN": 0.3,
+    "SWIFT_BIC": 0.3
+  },
+  "token_focal_gamma": 1.0,
+  "token_hard_fraction": 0.4,
+  "token_label_thresholds": {
+    "ACCOUNT_NUMBER": 0.18,
+    "BANK_ROUTING_NUMBER": 0.8,
+    "CREDIT_DEBIT_CARD": 0.8,
+    "EMAIL": 0.95,
+    "FIRST_NAME": 0.3,
+    "LAST_NAME": 0.4,
+    "PASSPORT_NUMBER": 0.75,
+    "PHONE_NUMBER": 0.65,
+    "POSTCODE": 0.9,
+    "PPSN": 0.7,
+    "SWIFT_BIC": 0.8
+  },
+  "token_positive_weight": 4.0,
+  "token_presence_weight": 1.0,
+  "transformers_version": "4.57.6",
+  "vocab_size": 119547
+}

eval/benchmark_summary.json

@@ -0,0 +1,66 @@
+{
+  "release": "IrishCore-DiffMask-135M-v1-rc6",
+  "repo_id": "temsa/IrishCore-DiffMask-135M-v1-rc6",
+  "architecture": {
+    "family": "DistilBERT-size token-span extractor",
+    "diffusion_style_training": true,
+    "runtime_diffusion": false,
+    "scanner_free": true,
+    "validator_free": true,
+    "heads": [
+      "token_presence_head",
+      "typed_start_boundary_head",
+      "typed_end_boundary_head"
+    ]
+  },
+  "base_model": "OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1",
+  "notes": [
+    "DiffMask uses a masked denoising training schedule, not a generative diffusion runtime.",
+    "ONNX q8 is the recommended CPU deployment artifact.",
+    "The release inference scripts emit [PII:LABEL] placeholders."
+  ],
+  "full": {
+    "core_f1": 0.9801324503311258,
+    "edge_f1": 0.9500000000000001,
+    "multilingual_f1": 0.9273743016759777,
+    "hardening_f1": 0.975609756097561
+  },
+  "onnx_q8": {
+    "core_f1": 0.9733333333333333,
+    "edge_f1": 0.9500000000000001,
+    "finance_f1": 0.9803921568627451,
+    "finance_boundary_f1": 1.0,
+    "user_ppsn_f1": 0.7499999999999999,
+    "gaelic_weak_ppsn_f1": 1.0,
+    "multilingual_f1": 0.9273743016759777,
+    "hardening_f1": 0.975609756097561,
+    "core_examples_per_second": 130.34154354965295,
+    "multilingual_examples_per_second": 200.20407351151675,
+    "runtime_profile_examples_per_second": 167.14610336175048
+  },
+  "comparison": {
+    "public_rc5_onnx_q8": {
+      "core": 0.9668874172185431,
+      "edge": 0.9743589743589743,
+      "remaining_gaps": 0.888888888888889,
+      "finance": 0.9361702127659575,
+      "finance_boundary": 0.8750000000000001,
+      "multilingual_ppsn": 0.9333333333333333,
+      "user_ppsn": 1.0,
+      "gaelic_weak_ppsn": 1.0,
+      "overlap_ppsn": 1.0
+    },
+    "public_rc8_onnx_q8": {
+      "min_score": 0.5,
+      "irish_core_manual_f1": 0.9736842105263158,
+      "irish_edge_f1": 1.0,
+      "finance_suite_f1": 1.0,
+      "finance_boundary_f1": 1.0,
+      "user_raw_ppsn_f1": 1.0,
+      "gaelic_weak_ppsn_f1": 1.0,
+      "multilingual_ppsn_f1": 0.9176470588235294,
+      "core_examples_per_second": 46.14201741375802,
+      "multilingual_examples_per_second": 99.71655616732895
+    }
+  }
+}

eval/benchmark_summary.md

@@ -0,0 +1,23 @@
+# Benchmark Summary
+
+## ONNX q8
+
+| Suite | F1 | Examples/s |
+|---|---:|---:|
+| Irish core | 0.9733 | 130.3415 |
+| Edge | 0.9500 | 275.6818 |
+| Finance | 0.9804 | 264.2625 |
+| Finance boundary | 1.0000 | 249.6232 |
+| User PPSN | 0.7500 | 146.1010 |
+| GA weak PPSN | 1.0000 | 67.8436 |
+| Multilingual PPSN | 0.9274 | 200.2041 |
+| Hardening exact | 0.9756 | 125.5843 |
+
+## Full checkpoint
+
+| Suite | F1 | Examples/s |
+|---|---:|---:|
+| Irish core | 0.9801 | 106.7506 |
+| Edge | 0.9500 | 114.6062 |
+| Multilingual PPSN | 0.9274 | 110.2363 |
+| Hardening exact | 0.9756 | 106.8223 |

eval/full_core_irish_core_pii_v1.json

@@ -0,0 +1,108 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30",
+  "input": "eval/irish_core_pii_v1.jsonl",
+  "loader_type": "token_span_pt",
+  "examples": 37,
+  "min_score": 0.5,
+  "iou_threshold": 0.5,
+  "elapsed_seconds": 0.3466022629290819,
+  "examples_per_second": 106.75060135879883,
+  "overall": {
+    "precision": 0.9866666666666667,
+    "recall": 0.9736842105263158,
+    "f1": 0.9801324503311258,
+    "tp": 74,
+    "fp": 1,
+    "fn": 2
+  },
+  "by_label": {
+    "ACCOUNT_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 3,
+      "fp": 0,
+      "fn": 0
+    },
+    "BANK_ROUTING_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 1,
+      "fp": 0,
+      "fn": 0
+    },
+    "CREDIT_DEBIT_CARD": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "EMAIL": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 6,
+      "fp": 0,
+      "fn": 0
+    },
+    "FIRST_NAME": {
+      "precision": 0.9444444444444444,
+      "recall": 0.8947368421052632,
+      "f1": 0.918918918918919,
+      "tp": 17,
+      "fp": 1,
+      "fn": 2
+    },
+    "LAST_NAME": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 19,
+      "fp": 0,
+      "fn": 0
+    },
+    "PASSPORT_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "PHONE_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 12,
+      "fp": 0,
+      "fn": 0
+    },
+    "POSTCODE": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 4,
+      "fp": 0,
+      "fn": 0
+    },
+    "PPSN": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 6,
+      "fp": 0,
+      "fn": 0
+    },
+    "SWIFT_BIC": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    }
+  }
+}

eval/full_edge_irish_ppsn_phone_edge_v1.json

@@ -0,0 +1,36 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30",
+  "input": "eval/irish_ppsn_phone_edge_v1.jsonl",
+  "loader_type": "token_span_pt",
+  "examples": 22,
+  "min_score": 0.5,
+  "iou_threshold": 0.5,
+  "elapsed_seconds": 0.19196176307741553,
+  "examples_per_second": 114.60615722271578,
+  "overall": {
+    "precision": 0.9047619047619048,
+    "recall": 1.0,
+    "f1": 0.9500000000000001,
+    "tp": 19,
+    "fp": 2,
+    "fn": 0
+  },
+  "by_label": {
+    "PHONE_NUMBER": {
+      "precision": 0.9285714285714286,
+      "recall": 1.0,
+      "f1": 0.962962962962963,
+      "tp": 13,
+      "fp": 1,
+      "fn": 0
+    },
+    "PPSN": {
+      "precision": 0.8571428571428571,
+      "recall": 1.0,
+      "f1": 0.923076923076923,
+      "tp": 6,
+      "fp": 1,
+      "fn": 0
+    }
+  }
+}

eval/full_hardening_exact_v1.json

@@ -0,0 +1,100 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30",
+  "input": "eval/irish_dllm_hardening_exact_v1.jsonl",
+  "loader_type": "token_span_pt",
+  "examples": 14,
+  "min_score": 0.5,
+  "iou_threshold": 0.5,
+  "elapsed_seconds": 0.13105878001078963,
+  "examples_per_second": 106.82229758927579,
+  "overall": {
+    "precision": 0.9523809523809523,
+    "recall": 1.0,
+    "f1": 0.975609756097561,
+    "tp": 20,
+    "fp": 1,
+    "fn": 0
+  },
+  "by_label": {
+    "ACCOUNT_NUMBER": {
+      "precision": 0.5,
+      "recall": 1.0,
+      "f1": 0.6666666666666666,
+      "tp": 1,
+      "fp": 1,
+      "fn": 0
+    },
+    "BANK_ROUTING_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "CREDIT_DEBIT_CARD": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "EMAIL": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "FIRST_NAME": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 1,
+      "fp": 0,
+      "fn": 0
+    },
+    "LAST_NAME": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 1,
+      "fp": 0,
+      "fn": 0
+    },
+    "PHONE_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "POSTCODE": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 3,
+      "fp": 0,
+      "fn": 0
+    },
+    "PPSN": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 3,
+      "fp": 0,
+      "fn": 0
+    },
+    "SWIFT_BIC": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 3,
+      "fp": 0,
+      "fn": 0
+    }
+  }
+}

eval/full_multilingual_ppsn_v1_all.json

@@ -0,0 +1,28 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30",
+  "input": "eval/multilingual_ppsn_v1_all.jsonl",
+  "loader_type": "token_span_pt",
+  "examples": 168,
+  "min_score": 0.5,
+  "iou_threshold": 0.5,
+  "elapsed_seconds": 1.5239994541043416,
+  "examples_per_second": 110.2362599589867,
+  "overall": {
+    "precision": 0.8736842105263158,
+    "recall": 0.9880952380952381,
+    "f1": 0.9273743016759777,
+    "tp": 83,
+    "fp": 12,
+    "fn": 1
+  },
+  "by_label": {
+    "PPSN": {
+      "precision": 0.9325842696629213,
+      "recall": 0.9880952380952381,
+      "f1": 0.9595375722543353,
+      "tp": 83,
+      "fp": 6,
+      "fn": 1
+    }
+  }
+}

eval/harness_reconciliation.json

@@ -0,0 +1,109 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6b-focusv10-e012-b48w0",
+  "datasets": {
+    "fresh_holdout": {
+      "benchmark_multitask_pt": {
+        "precision": 0.9047619047619048,
+        "recall": 0.59375,
+        "f1": 0.7169811320754718,
+        "tp": 19,
+        "fp": 2,
+        "fn": 13
+      },
+      "eval_release_clean_single_pass": {
+        "precision": 0.9047619047619048,
+        "recall": 0.59375,
+        "f1": 0.7169811320754718,
+        "tp": 19,
+        "fp": 2,
+        "fn": 13
+      },
+      "eval_release_diffusion_last_pass": {
+        "precision": 0.782608695652174,
+        "recall": 0.5625,
+        "f1": 0.6545454545454547,
+        "tp": 18,
+        "fp": 5,
+        "fn": 14
+      }
+    },
+    "uat_exact": {
+      "benchmark_multitask_pt": {
+        "precision": 1.0,
+        "recall": 0.8235294117647058,
+        "f1": 0.9032258064516129,
+        "tp": 14,
+        "fp": 0,
+        "fn": 3
+      },
+      "eval_release_clean_single_pass": {
+        "precision": 1.0,
+        "recall": 0.8235294117647058,
+        "f1": 0.9032258064516129,
+        "tp": 14,
+        "fp": 0,
+        "fn": 3
+      },
+      "eval_release_diffusion_last_pass": {
+        "precision": 1.0,
+        "recall": 0.8235294117647058,
+        "f1": 0.9032258064516129,
+        "tp": 14,
+        "fp": 0,
+        "fn": 3
+      }
+    },
+    "irish_core": {
+      "benchmark_multitask_pt": {
+        "precision": 0.9864864864864865,
+        "recall": 0.9605263157894737,
+        "f1": 0.9733333333333333,
+        "tp": 73,
+        "fp": 1,
+        "fn": 3
+      },
+      "eval_release_clean_single_pass": {
+        "precision": 0.9864864864864865,
+        "recall": 0.9605263157894737,
+        "f1": 0.9733333333333333,
+        "tp": 73,
+        "fp": 1,
+        "fn": 3
+      },
+      "eval_release_diffusion_last_pass": {
+        "precision": 0.9736842105263158,
+        "recall": 0.9736842105263158,
+        "f1": 0.9736842105263158,
+        "tp": 74,
+        "fp": 2,
+        "fn": 2
+      }
+    },
+    "multilingual_ppsn": {
+      "benchmark_multitask_pt": {
+        "precision": 0.8736842105263158,
+        "recall": 0.9880952380952381,
+        "f1": 0.9273743016759777,
+        "tp": 83,
+        "fp": 12,
+        "fn": 1
+      },
+      "eval_release_clean_single_pass": {
+        "precision": 0.8736842105263158,
+        "recall": 0.9880952380952381,
+        "f1": 0.9273743016759777,
+        "tp": 83,
+        "fp": 12,
+        "fn": 1
+      },
+      "eval_release_diffusion_last_pass": {
+        "precision": 0.8666666666666667,
+        "recall": 0.9285714285714286,
+        "f1": 0.896551724137931,
+        "tp": 78,
+        "fp": 12,
+        "fn": 6
+      }
+    }
+  }
+}

eval/harness_reconciliation.md

@@ -0,0 +1,24 @@
+# DiffMask Eval Harness
+
+- Model: `models/irishcore-diffmask-135m-v1-rc6b-focusv10-e012-b48w0`
+
+## Result
+
+The deployment-aligned harness is the clean single-pass token-span path.
+
+- `experiments/irish_core_span_raw_only/benchmark_multitask.py` and `scripts/eval_dllm_release.py --inference-mode clean_single_pass` match exactly on the checked suites.
+- The old diffusion-style eval path (`diffusion_last_pass`) is not deployment-aligned and depresses scores on several suites.
+
+## Comparison
+
+| Dataset | benchmark_multitask | eval_clean_single_pass | eval_diffusion_last_pass |
+|---|---:|---:|---:|
+| `fresh_holdout` | 0.7170 | 0.7170 | 0.6545 |
+| `uat_exact` | 0.9032 | 0.9032 | 0.9032 |
+| `irish_core` | 0.9733 | 0.9733 | 0.9737 |
+| `multilingual_ppsn` | 0.9274 | 0.9274 | 0.8966 |
+
+## Conclusion
+
+- Use `clean_single_pass` for release gating and model comparison.
+- Keep `diffusion_last_pass` only as a training diagnostic if needed.

eval/q8_core_irish_core_pii_v1.json

@@ -0,0 +1,108 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30_onnx_q8",
+  "input": "eval/irish_core_pii_v1.jsonl",
+  "loader_type": "token_span_onnx_q8",
+  "examples": 37,
+  "min_score": 0.5,
+  "iou_threshold": 0.5,
+  "elapsed_seconds": 0.28386958595365286,
+  "examples_per_second": 130.34154354965295,
+  "overall": {
+    "precision": 0.9864864864864865,
+    "recall": 0.9605263157894737,
+    "f1": 0.9733333333333333,
+    "tp": 73,
+    "fp": 1,
+    "fn": 3
+  },
+  "by_label": {
+    "ACCOUNT_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 3,
+      "fp": 0,
+      "fn": 0
+    },
+    "BANK_ROUTING_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 1,
+      "fp": 0,
+      "fn": 0
+    },
+    "CREDIT_DEBIT_CARD": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "EMAIL": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 6,
+      "fp": 0,
+      "fn": 0
+    },
+    "FIRST_NAME": {
+      "precision": 0.9444444444444444,
+      "recall": 0.8947368421052632,
+      "f1": 0.918918918918919,
+      "tp": 17,
+      "fp": 1,
+      "fn": 2
+    },
+    "LAST_NAME": {
+      "precision": 1.0,
+      "recall": 0.9473684210526315,
+      "f1": 0.972972972972973,
+      "tp": 18,
+      "fp": 0,
+      "fn": 1
+    },
+    "PASSPORT_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "PHONE_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 12,
+      "fp": 0,
+      "fn": 0
+    },
+    "POSTCODE": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 4,
+      "fp": 0,
+      "fn": 0
+    },
+    "PPSN": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 6,
+      "fp": 0,
+      "fn": 0
+    },
+    "SWIFT_BIC": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    }
+  }
+}

eval/q8_edge_irish_ppsn_phone_edge_v1.json

@@ -0,0 +1,36 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30_onnx_q8",
+  "input": "eval/irish_ppsn_phone_edge_v1.jsonl",
+  "loader_type": "token_span_onnx_q8",
+  "examples": 22,
+  "min_score": 0.5,
+  "iou_threshold": 0.5,
+  "elapsed_seconds": 0.0798021589871496,
+  "examples_per_second": 275.68176449389824,
+  "overall": {
+    "precision": 0.9047619047619048,
+    "recall": 1.0,
+    "f1": 0.9500000000000001,
+    "tp": 19,
+    "fp": 2,
+    "fn": 0
+  },
+  "by_label": {
+    "PHONE_NUMBER": {
+      "precision": 0.9285714285714286,
+      "recall": 1.0,
+      "f1": 0.962962962962963,
+      "tp": 13,
+      "fp": 1,
+      "fn": 0
+    },
+    "PPSN": {
+      "precision": 0.8571428571428571,
+      "recall": 1.0,
+      "f1": 0.923076923076923,
+      "tp": 6,
+      "fp": 1,
+      "fn": 0
+    }
+  }
+}

eval/q8_finance_boundary_repair_v1.json

@@ -0,0 +1,76 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30_onnx_q8",
+  "input": "eval/irish_finance_boundary_repair_v1.jsonl",
+  "loader_type": "token_span_onnx_q8",
+  "examples": 12,
+  "min_score": 0.5,
+  "iou_threshold": 0.5,
+  "elapsed_seconds": 0.0480724589433521,
+  "examples_per_second": 249.62317850519418,
+  "overall": {
+    "precision": 1.0,
+    "recall": 1.0,
+    "f1": 1.0,
+    "tp": 18,
+    "fp": 0,
+    "fn": 0
+  },
+  "by_label": {
+    "ACCOUNT_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "BANK_ROUTING_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "CREDIT_DEBIT_CARD": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "PASSPORT_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 4,
+      "fp": 0,
+      "fn": 0
+    },
+    "PHONE_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 4,
+      "fp": 0,
+      "fn": 0
+    },
+    "PPSN": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "SWIFT_BIC": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    }
+  }
+}

eval/q8_finance_irish_phone_passport_finance_v1.json

@@ -0,0 +1,76 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30_onnx_q8",
+  "input": "eval/irish_phone_passport_finance_v1.jsonl",
+  "loader_type": "token_span_onnx_q8",
+  "examples": 20,
+  "min_score": 0.5,
+  "iou_threshold": 0.5,
+  "elapsed_seconds": 0.07568231993354857,
+  "examples_per_second": 264.2625122691881,
+  "overall": {
+    "precision": 0.9615384615384616,
+    "recall": 1.0,
+    "f1": 0.9803921568627451,
+    "tp": 25,
+    "fp": 1,
+    "fn": 0
+  },
+  "by_label": {
+    "ACCOUNT_NUMBER": {
+      "precision": 0.6666666666666666,
+      "recall": 1.0,
+      "f1": 0.8,
+      "tp": 2,
+      "fp": 1,
+      "fn": 0
+    },
+    "BANK_ROUTING_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 5,
+      "fp": 0,
+      "fn": 0
+    },
+    "CREDIT_DEBIT_CARD": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "PASSPORT_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 6,
+      "fp": 0,
+      "fn": 0
+    },
+    "PHONE_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 6,
+      "fp": 0,
+      "fn": 0
+    },
+    "PPSN": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "SWIFT_BIC": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    }
+  }
+}

eval/q8_gaelic_weak_ppsn_v1.json

@@ -0,0 +1,28 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30_onnx_q8",
+  "input": "eval/qa_feedback_ga_ppsn_weakctx_v1.jsonl",
+  "loader_type": "token_span_onnx_q8",
+  "examples": 2,
+  "min_score": 0.5,
+  "iou_threshold": 0.5,
+  "elapsed_seconds": 0.02947957592550665,
+  "examples_per_second": 67.84358109675307,
+  "overall": {
+    "precision": 1.0,
+    "recall": 1.0,
+    "f1": 1.0,
+    "tp": 2,
+    "fp": 0,
+    "fn": 0
+  },
+  "by_label": {
+    "PPSN": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    }
+  }
+}

eval/q8_hardening_exact_v1.json

@@ -0,0 +1,100 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30_onnx_q8",
+  "input": "eval/irish_dllm_hardening_exact_v1.jsonl",
+  "loader_type": "token_span_onnx_q8",
+  "examples": 14,
+  "min_score": 0.5,
+  "iou_threshold": 0.5,
+  "elapsed_seconds": 0.11147892102599144,
+  "examples_per_second": 125.58427971092296,
+  "overall": {
+    "precision": 0.9523809523809523,
+    "recall": 1.0,
+    "f1": 0.975609756097561,
+    "tp": 20,
+    "fp": 1,
+    "fn": 0
+  },
+  "by_label": {
+    "ACCOUNT_NUMBER": {
+      "precision": 0.5,
+      "recall": 1.0,
+      "f1": 0.6666666666666666,
+      "tp": 1,
+      "fp": 1,
+      "fn": 0
+    },
+    "BANK_ROUTING_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "CREDIT_DEBIT_CARD": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "EMAIL": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "FIRST_NAME": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 1,
+      "fp": 0,
+      "fn": 0
+    },
+    "LAST_NAME": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 1,
+      "fp": 0,
+      "fn": 0
+    },
+    "PHONE_NUMBER": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 2,
+      "fp": 0,
+      "fn": 0
+    },
+    "POSTCODE": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 3,
+      "fp": 0,
+      "fn": 0
+    },
+    "PPSN": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 3,
+      "fp": 0,
+      "fn": 0
+    },
+    "SWIFT_BIC": {
+      "precision": 1.0,
+      "recall": 1.0,
+      "f1": 1.0,
+      "tp": 3,
+      "fp": 0,
+      "fn": 0
+    }
+  }
+}

eval/q8_multilingual_ppsn_v1_all.json

@@ -0,0 +1,28 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30_onnx_q8",
+  "input": "eval/multilingual_ppsn_v1_all.jsonl",
+  "loader_type": "token_span_onnx_q8",
+  "examples": 168,
+  "min_score": 0.5,
+  "iou_threshold": 0.5,
+  "elapsed_seconds": 0.8391437649261206,
+  "examples_per_second": 200.20407351151675,
+  "overall": {
+    "precision": 0.8736842105263158,
+    "recall": 0.9880952380952381,
+    "f1": 0.9273743016759777,
+    "tp": 83,
+    "fp": 12,
+    "fn": 1
+  },
+  "by_label": {
+    "PPSN": {
+      "precision": 0.9325842696629213,
+      "recall": 0.9880952380952381,
+      "f1": 0.9595375722543353,
+      "tp": 83,
+      "fp": 6,
+      "fn": 1
+    }
+  }
+}

eval/q8_runtime_profile_source.json

@@ -0,0 +1,28 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30_onnx_q8",
+  "input": "eval/runtime_profile_source.jsonl",
+  "loader_type": "token_span_onnx_q8",
+  "examples": 3223,
+  "min_score": 0.5,
+  "iou_threshold": 0.5,
+  "elapsed_seconds": 19.28253148100339,
+  "examples_per_second": 167.14610336175048,
+  "overall": {
+    "precision": 0.5172647257955315,
+    "recall": 0.21539329010431352,
+    "f1": 0.304140127388535,
+    "tp": 764,
+    "fp": 713,
+    "fn": 2783
+  },
+  "by_label": {
+    "PPSN": {
+      "precision": 0.7900723888314375,
+      "recall": 0.21539329010431352,
+      "f1": 0.3385024368630926,
+      "tp": 764,
+      "fp": 203,
+      "fn": 2783
+    }
+  }
+}

eval/q8_user_raw_regression_cases_v1.json

@@ -0,0 +1,28 @@
+{
+  "model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30_onnx_q8",
+  "input": "eval/user_raw_regression_cases_v1.jsonl",
+  "loader_type": "token_span_onnx_q8",
+  "examples": 7,
+  "min_score": 0.5,
+  "iou_threshold": 0.5,
+  "elapsed_seconds": 0.04791205294895917,
+  "examples_per_second": 146.1010240462273,
+  "overall": {
+    "precision": 0.6,
+    "recall": 1.0,
+    "f1": 0.7499999999999999,
+    "tp": 3,
+    "fp": 2,
+    "fn": 0
+  },
+  "by_label": {
+    "PPSN": {
+      "precision": 0.75,
+      "recall": 1.0,
+      "f1": 0.8571428571428571,
+      "tp": 3,
+      "fp": 1,
+      "fn": 0
+    }
+  }
+}

inference_mask.py

@@ -0,0 +1,111 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import json
+import os
+
+os.environ.setdefault("TRANSFORMERS_NO_TF", "1")
+os.environ.setdefault("TRANSFORMERS_NO_FLAX", "1")
+os.environ.setdefault("TRANSFORMERS_NO_TORCHVISION", "1")
+os.environ["USE_TF"] = "0"
+os.environ["USE_FLAX"] = "0"
+os.environ["USE_TORCH"] = "1"
+
+import torch
+from transformers import AutoConfig
+
+from common import (
+    boundary_label_thresholds_from_config,
+    decode_token_presence_segments,
+    label_max_span_tokens_from_config,
+    label_min_nonspace_chars_from_config,
+    label_names_from_config,
+    safe_auto_tokenizer,
+    token_extend_thresholds_from_config,
+    token_label_thresholds_from_config,
+)
+from multitask_model import IrishCoreTokenSpanModel
+
+
+def replacement(label: str) -> str:
+    return f"[PII:{label}]"
+
+
+def mask_text(text: str, spans: list[dict]) -> str:
+    out = text
+    for span in sorted(spans, key=lambda item: (item["start"], item["end"]), reverse=True):
+        out = out[: span["start"]] + replacement(span["label"]) + out[span["end"] :]
+    return out
+
+
+def predict(text: str, model, tokenizer, min_score: float):
+    encoded = tokenizer(text, return_offsets_mapping=True, return_tensors="pt", truncation=True)
+    offsets = [tuple(item) for item in encoded.pop("offset_mapping")[0].tolist()]
+    device = next(model.parameters()).device
+    encoded = {key: value.to(device) for key, value in encoded.items()}
+    with torch.no_grad():
+        output = model(**encoded)
+    token_scores = torch.sigmoid(output.token_logits[0]).cpu().numpy()
+    start_scores = torch.sigmoid(output.start_logits[0]).cpu().numpy()
+    end_scores = torch.sigmoid(output.end_logits[0]).cpu().numpy()
+    label_names = label_names_from_config(model.config)
+    thresholds = token_label_thresholds_from_config(model.config, min_score)
+    extend_thresholds = token_extend_thresholds_from_config(model.config)
+    max_span_tokens = label_max_span_tokens_from_config(model.config)
+    min_nonspace_chars = label_min_nonspace_chars_from_config(model.config)
+    boundary_thresholds = boundary_label_thresholds_from_config(model.config)
+    spans = decode_token_presence_segments(
+        text,
+        offsets,
+        token_scores,
+        label_names,
+        min_score,
+        thresholds,
+        extend_thresholds,
+        max_span_tokens,
+        min_nonspace_chars,
+        boundary_thresholds,
+        start_scores=start_scores,
+        end_scores=end_scores,
+    )
+    for span in spans:
+        span["replacement"] = replacement(span["label"])
+    return spans
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", required=True)
+    parser.add_argument("--text", required=True)
+    parser.add_argument("--min-score", type=float, default=0.5)
+    parser.add_argument("--device", choices=["auto", "cpu", "cuda"], default="auto")
+    parser.add_argument("--json", action="store_true")
+    args = parser.parse_args()
+
+    tokenizer = safe_auto_tokenizer(args.model)
+    config = AutoConfig.from_pretrained(args.model)
+    model = IrishCoreTokenSpanModel.from_pretrained(args.model, config=config)
+    if args.device == "auto":
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+    else:
+        device = args.device
+    model.to(device)
+    model.eval()
+
+    spans = predict(args.text, model, tokenizer, args.min_score)
+    result = {
+        "model": args.model,
+        "backend": "transformers_token_span",
+        "min_score": args.min_score,
+        "spans": spans,
+        "masked_text": mask_text(args.text, spans),
+    }
+    if args.json:
+        print(json.dumps(result, indent=2, ensure_ascii=False))
+    else:
+        print(result["masked_text"])
+
+
+if __name__ == "__main__":
+    main()

inference_mask_onnx.py

@@ -0,0 +1,97 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import json
+import os
+
+os.environ.setdefault("TRANSFORMERS_NO_TF", "1")
+os.environ.setdefault("TRANSFORMERS_NO_FLAX", "1")
+os.environ.setdefault("TRANSFORMERS_NO_TORCHVISION", "1")
+os.environ["USE_TF"] = "0"
+os.environ["USE_FLAX"] = "0"
+os.environ["USE_TORCH"] = "1"
+
+import numpy as np
+
+from common import (
+    boundary_label_thresholds_from_config,
+    decode_token_presence_segments,
+    label_max_span_tokens_from_config,
+    label_min_nonspace_chars_from_config,
+    label_names_from_config,
+    load_onnx_session,
+    run_onnx_all,
+    token_extend_thresholds_from_config,
+    token_label_thresholds_from_config,
+)
+
+
+def replacement(label: str) -> str:
+    return f"[PII:{label}]"
+
+
+def mask_text(text: str, spans: list[dict]) -> str:
+    out = text
+    for span in sorted(spans, key=lambda item: (item["start"], item["end"]), reverse=True):
+        out = out[: span["start"]] + replacement(span["label"]) + out[span["end"] :]
+    return out
+
+
+def predict(text: str, session, tokenizer, config, min_score: float):
+    encoded = tokenizer(text, return_offsets_mapping=True, return_tensors="np", truncation=True)
+    offsets = [tuple(item) for item in encoded["offset_mapping"][0].tolist()]
+    token_logits, start_logits, end_logits = run_onnx_all(session, encoded)
+    token_scores = 1.0 / (1.0 + np.exp(-token_logits[0]))
+    start_scores = 1.0 / (1.0 + np.exp(-start_logits[0]))
+    end_scores = 1.0 / (1.0 + np.exp(-end_logits[0]))
+    label_names = label_names_from_config(config)
+    thresholds = token_label_thresholds_from_config(config, min_score)
+    extend_thresholds = token_extend_thresholds_from_config(config)
+    max_span_tokens = label_max_span_tokens_from_config(config)
+    min_nonspace_chars = label_min_nonspace_chars_from_config(config)
+    boundary_thresholds = boundary_label_thresholds_from_config(config)
+    spans = decode_token_presence_segments(
+        text,
+        offsets,
+        token_scores,
+        label_names,
+        min_score,
+        thresholds,
+        extend_thresholds,
+        max_span_tokens,
+        min_nonspace_chars,
+        boundary_thresholds,
+        start_scores=start_scores,
+        end_scores=end_scores,
+    )
+    for span in spans:
+        span["replacement"] = replacement(span["label"])
+    return spans
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", required=True)
+    parser.add_argument("--text", required=True)
+    parser.add_argument("--min-score", type=float, default=0.5)
+    parser.add_argument("--json", action="store_true")
+    args = parser.parse_args()
+
+    session, tokenizer, config = load_onnx_session(args.model, onnx_file="model_quantized.onnx", onnx_subfolder="onnx")
+    spans = predict(args.text, session, tokenizer, config, args.min_score)
+    result = {
+        "model": args.model,
+        "backend": "onnx_token_span_q8",
+        "min_score": args.min_score,
+        "spans": spans,
+        "masked_text": mask_text(args.text, spans),
+    }
+    if args.json:
+        print(json.dumps(result, indent=2, ensure_ascii=False))
+    else:
+        print(result["masked_text"])
+
+
+if __name__ == "__main__":
+    main()

model.py

@@ -0,0 +1,75 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+import torch.nn as nn
+from transformers import AutoConfig, AutoModel, PreTrainedModel
+from transformers.utils import ModelOutput
+
+
+def hidden_size_from_config(config) -> int:
+    return int(getattr(config, "hidden_size", getattr(config, "dim")))
+
+
+@dataclass
+class MultilabelSpanOutput(ModelOutput):
+    loss: Optional[torch.Tensor] = None
+    start_logits: Optional[torch.Tensor] = None
+    end_logits: Optional[torch.Tensor] = None
+
+
+class IrishCoreSpanHeadModel(PreTrainedModel):
+    config_class = AutoConfig
+    base_model_prefix = "encoder"
+
+    def __init__(self, config):
+        super().__init__(config)
+        num_span_labels = int(getattr(config, "num_span_labels"))
+        self.encoder = AutoModel.from_config(config)
+        hidden_size = hidden_size_from_config(config)
+        dropout = float(getattr(config, "seq_classif_dropout", getattr(config, "dropout", 0.1)))
+        self.dropout = nn.Dropout(dropout)
+        self.start_classifier = nn.Linear(hidden_size, num_span_labels)
+        self.end_classifier = nn.Linear(hidden_size, num_span_labels)
+        pos_weight = float(getattr(config, "span_positive_weight", 6.0))
+        self.register_buffer("loss_pos_weight", torch.full((num_span_labels,), pos_weight), persistent=False)
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        start_positions=None,
+        end_positions=None,
+        token_mask=None,
+        **kwargs,
+    ) -> MultilabelSpanOutput:
+        encoder_kwargs = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            **kwargs,
+        }
+        if token_type_ids is not None and getattr(self.config, "model_type", "") not in {"distilbert", "roberta"}:
+            encoder_kwargs["token_type_ids"] = token_type_ids
+        outputs = self.encoder(**encoder_kwargs)
+        hidden = self.dropout(outputs.last_hidden_state)
+        start_logits = self.start_classifier(hidden)
+        end_logits = self.end_classifier(hidden)
+
+        loss = None
+        if start_positions is not None and end_positions is not None:
+            if token_mask is None:
+                token_mask = attention_mask
+            mask = token_mask.float().unsqueeze(-1)
+            pos_weight = self.loss_pos_weight.to(start_logits.device)
+            bce = nn.BCEWithLogitsLoss(reduction="none", pos_weight=pos_weight)
+            start_loss = bce(start_logits, start_positions.float()) * mask
+            end_loss = bce(end_logits, end_positions.float()) * mask
+            denom = mask.sum().clamp_min(1.0) * start_logits.shape[-1]
+            loss = (start_loss.sum() + end_loss.sum()) / (2.0 * denom)
+
+        return MultilabelSpanOutput(loss=loss, start_logits=start_logits, end_logits=end_logits)

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50d5c48025ce05955cce8a480ce09b1051e55a363a58d73b124b89ff90774688
+size 539050252

multitask_model.py

@@ -0,0 +1,125 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+from dataclasses import dataclass
+import math
+from typing import Optional
+
+import torch
+import torch.nn as nn
+from transformers import AutoConfig, AutoModel, PreTrainedModel
+from transformers.utils import ModelOutput
+
+try:
+    from .model import hidden_size_from_config
+except ImportError:
+    from model import hidden_size_from_config
+
+
+@dataclass
+class MultitaskSpanOutput(ModelOutput):
+    loss: Optional[torch.Tensor] = None
+    token_logits: Optional[torch.Tensor] = None
+    start_logits: Optional[torch.Tensor] = None
+    end_logits: Optional[torch.Tensor] = None
+
+
+class IrishCoreTokenSpanModel(PreTrainedModel):
+    config_class = AutoConfig
+    base_model_prefix = "encoder"
+
+    def __init__(self, config):
+        super().__init__(config)
+        num_span_labels = int(getattr(config, "num_span_labels"))
+        self.encoder = AutoModel.from_config(config)
+        hidden_size = hidden_size_from_config(config)
+        dropout = float(getattr(config, "seq_classif_dropout", getattr(config, "dropout", 0.1)))
+        self.dropout = nn.Dropout(dropout)
+        self.token_classifier = nn.Linear(hidden_size, num_span_labels)
+        self.start_classifier = nn.Linear(hidden_size, num_span_labels)
+        self.end_classifier = nn.Linear(hidden_size, num_span_labels)
+        boundary_pos_weight = float(getattr(config, "span_positive_weight", 6.0))
+        presence_pos_weight = float(getattr(config, "token_positive_weight", 4.0))
+        self.register_buffer("boundary_pos_weight", torch.full((num_span_labels,), boundary_pos_weight), persistent=False)
+        self.register_buffer("presence_pos_weight", torch.full((num_span_labels,), presence_pos_weight), persistent=False)
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        token_labels=None,
+        start_positions=None,
+        end_positions=None,
+        token_mask=None,
+        **kwargs,
+    ) -> MultitaskSpanOutput:
+        encoder_kwargs = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            **kwargs,
+        }
+        if token_type_ids is not None and getattr(self.config, "model_type", "") not in {"distilbert", "roberta"}:
+            encoder_kwargs["token_type_ids"] = token_type_ids
+        outputs = self.encoder(**encoder_kwargs)
+        hidden = self.dropout(outputs.last_hidden_state)
+        token_logits = self.token_classifier(hidden)
+        start_logits = self.start_classifier(hidden)
+        end_logits = self.end_classifier(hidden)
+
+        loss = None
+        if token_labels is not None and start_positions is not None and end_positions is not None:
+            if token_mask is None:
+                token_mask = attention_mask
+            mask = token_mask.float().unsqueeze(-1)
+            boundary_pos_weight = self.boundary_pos_weight.to(token_logits.device)
+            presence_pos_weight = self.presence_pos_weight.to(token_logits.device)
+            bce_boundary = nn.BCEWithLogitsLoss(reduction="none", pos_weight=boundary_pos_weight)
+            bce_presence = nn.BCEWithLogitsLoss(reduction="none", pos_weight=presence_pos_weight)
+            token_targets = token_labels.float()
+            start_targets = start_positions.float()
+            end_targets = end_positions.float()
+            token_loss = bce_presence(token_logits, token_targets)
+            start_loss = bce_boundary(start_logits, start_targets)
+            end_loss = bce_boundary(end_logits, end_targets)
+            token_focal_gamma = float(getattr(self.config, "token_focal_gamma", getattr(self.config, "focal_gamma", 0.0)))
+            boundary_focal_gamma = float(getattr(self.config, "boundary_focal_gamma", getattr(self.config, "focal_gamma", 0.0)))
+            if token_focal_gamma > 0.0:
+                token_loss = apply_focal_weight(token_loss, token_logits, token_targets, token_focal_gamma)
+            if boundary_focal_gamma > 0.0:
+                start_loss = apply_focal_weight(start_loss, start_logits, start_targets, boundary_focal_gamma)
+                end_loss = apply_focal_weight(end_loss, end_logits, end_targets, boundary_focal_gamma)
+            token_hard_fraction = float(getattr(self.config, "token_hard_fraction", getattr(self.config, "hard_fraction", 1.0)))
+            boundary_hard_fraction = float(getattr(self.config, "boundary_hard_fraction", getattr(self.config, "hard_fraction", 1.0)))
+            token_loss = reduce_masked_loss(token_loss, mask, token_hard_fraction)
+            start_loss = reduce_masked_loss(start_loss, mask, boundary_hard_fraction)
+            end_loss = reduce_masked_loss(end_loss, mask, boundary_hard_fraction)
+            boundary_loss = 0.5 * (start_loss + end_loss)
+            token_weight = float(getattr(self.config, "token_presence_weight", 1.0))
+            boundary_weight = float(getattr(self.config, "boundary_loss_weight", 1.0))
+            loss = token_weight * token_loss + boundary_weight * boundary_loss
+
+        return MultitaskSpanOutput(
+            loss=loss,
+            token_logits=token_logits,
+            start_logits=start_logits,
+            end_logits=end_logits,
+        )
+
+
+def apply_focal_weight(loss: torch.Tensor, logits: torch.Tensor, targets: torch.Tensor, gamma: float) -> torch.Tensor:
+    probs = torch.sigmoid(logits)
+    pt = probs * targets + (1.0 - probs) * (1.0 - targets)
+    return loss * (1.0 - pt).pow(gamma)
+
+
+def reduce_masked_loss(loss: torch.Tensor, mask: torch.Tensor, hard_fraction: float) -> torch.Tensor:
+    expanded_mask = mask.expand_as(loss).bool()
+    masked = loss.masked_select(expanded_mask)
+    if masked.numel() == 0:
+        return loss.sum() * 0.0
+    if 0.0 < hard_fraction < 1.0:
+        keep = max(1, math.ceil(masked.numel() * hard_fraction))
+        masked = torch.topk(masked, keep).values
+    return masked.mean()

onnx/config.json

@@ -0,0 +1,338 @@
+{
+  "activation": "gelu",
+  "architectures": [
+    "IrishCoreTokenSpanModel"
+  ],
+  "attention_dropout": 0.1,
+  "boundary_focal_gamma": 2.0,
+  "boundary_hard_fraction": 0.25,
+  "boundary_label_thresholds": {
+    "ACCOUNT_NUMBER": 0.1,
+    "BANK_ROUTING_NUMBER": 0.25,
+    "CREDIT_DEBIT_CARD": 0.25,
+    "EMAIL": 0.2,
+    "FIRST_NAME": 0.1,
+    "LAST_NAME": 0.1,
+    "PASSPORT_NUMBER": 0.25,
+    "PHONE_NUMBER": 0.25,
+    "POSTCODE": 0.4,
+    "PPSN": 0.35,
+    "SWIFT_BIC": 0.25
+  },
+  "boundary_loss_weight": 1.0,
+  "decoder_profile": "rc3_focusv3_email_extend015_passport075",
+  "dim": 768,
+  "dropout": 0.1,
+  "dtype": "float32",
+  "hidden_dim": 3072,
+  "id2label": {
+    "0": "O",
+    "1": "B-account_number",
+    "2": "B-age",
+    "3": "B-api_key",
+    "4": "B-bank_routing_number",
+    "5": "B-biometric_identifier",
+    "6": "B-blood_type",
+    "7": "B-certificate_license_number",
+    "8": "B-city",
+    "9": "B-company_name",
+    "10": "B-coordinate",
+    "11": "B-country",
+    "12": "B-county",
+    "13": "B-credit_debit_card",
+    "14": "B-customer_id",
+    "15": "B-cvv",
+    "16": "B-date",
+    "17": "B-date_of_birth",
+    "18": "B-date_time",
+    "19": "B-device_identifier",
+    "20": "B-education_level",
+    "21": "B-email",
+    "22": "B-employee_id",
+    "23": "B-employment_status",
+    "24": "B-fax_number",
+    "25": "B-first_name",
+    "26": "B-gender",
+    "27": "B-health_plan_beneficiary_number",
+    "28": "B-http_cookie",
+    "29": "B-ipv4",
+    "30": "B-ipv6",
+    "31": "B-language",
+    "32": "B-last_name",
+    "33": "B-license_plate",
+    "34": "B-mac_address",
+    "35": "B-medical_record_number",
+    "36": "B-occupation",
+    "37": "B-password",
+    "38": "B-phone_number",
+    "39": "B-pin",
+    "40": "B-political_view",
+    "41": "B-postcode",
+    "42": "B-race_ethnicity",
+    "43": "B-religious_belief",
+    "44": "B-sexuality",
+    "45": "B-ssn",
+    "46": "B-state",
+    "47": "B-street_address",
+    "48": "B-swift_bic",
+    "49": "B-tax_id",
+    "50": "B-time",
+    "51": "B-unique_id",
+    "52": "B-url",
+    "53": "B-user_name",
+    "54": "B-vehicle_identifier",
+    "55": "I-account_number",
+    "56": "I-api_key",
+    "57": "I-biometric_identifier",
+    "58": "I-blood_type",
+    "59": "I-certificate_license_number",
+    "60": "I-city",
+    "61": "I-company_name",
+    "62": "I-coordinate",
+    "63": "I-country",
+    "64": "I-county",
+    "65": "I-credit_debit_card",
+    "66": "I-customer_id",
+    "67": "I-date",
+    "68": "I-date_of_birth",
+    "69": "I-date_time",
+    "70": "I-device_identifier",
+    "71": "I-education_level",
+    "72": "I-email",
+    "73": "I-employee_id",
+    "74": "I-employment_status",
+    "75": "I-fax_number",
+    "76": "I-first_name",
+    "77": "I-gender",
+    "78": "I-health_plan_beneficiary_number",
+    "79": "I-http_cookie",
+    "80": "I-ipv4",
+    "81": "I-ipv6",
+    "82": "I-language",
+    "83": "I-last_name",
+    "84": "I-license_plate",
+    "85": "I-mac_address",
+    "86": "I-medical_record_number",
+    "87": "I-occupation",
+    "88": "I-password",
+    "89": "I-phone_number",
+    "90": "I-pin",
+    "91": "I-political_view",
+    "92": "I-postcode",
+    "93": "I-race_ethnicity",
+    "94": "I-religious_belief",
+    "95": "I-sexuality",
+    "96": "I-ssn",
+    "97": "I-state",
+    "98": "I-street_address",
+    "99": "I-swift_bic",
+    "100": "I-tax_id",
+    "101": "I-time",
+    "102": "I-unique_id",
+    "103": "I-url",
+    "104": "I-user_name",
+    "105": "I-vehicle_identifier",
+    "106": "B-PPSN",
+    "107": "I-PPSN",
+    "108": "B-PASSPORT_NUMBER",
+    "109": "I-PASSPORT_NUMBER",
+    "110": "I-bank_routing_number"
+  },
+  "initializer_range": 0.02,
+  "label2id": {
+    "B-PASSPORT_NUMBER": 108,
+    "B-PPSN": 106,
+    "B-account_number": 1,
+    "B-age": 2,
+    "B-api_key": 3,
+    "B-bank_routing_number": 4,
+    "B-biometric_identifier": 5,
+    "B-blood_type": 6,
+    "B-certificate_license_number": 7,
+    "B-city": 8,
+    "B-company_name": 9,
+    "B-coordinate": 10,
+    "B-country": 11,
+    "B-county": 12,
+    "B-credit_debit_card": 13,
+    "B-customer_id": 14,
+    "B-cvv": 15,
+    "B-date": 16,
+    "B-date_of_birth": 17,
+    "B-date_time": 18,
+    "B-device_identifier": 19,
+    "B-education_level": 20,
+    "B-email": 21,
+    "B-employee_id": 22,
+    "B-employment_status": 23,
+    "B-fax_number": 24,
+    "B-first_name": 25,
+    "B-gender": 26,
+    "B-health_plan_beneficiary_number": 27,
+    "B-http_cookie": 28,
+    "B-ipv4": 29,
+    "B-ipv6": 30,
+    "B-language": 31,
+    "B-last_name": 32,
+    "B-license_plate": 33,
+    "B-mac_address": 34,
+    "B-medical_record_number": 35,
+    "B-occupation": 36,
+    "B-password": 37,
+    "B-phone_number": 38,
+    "B-pin": 39,
+    "B-political_view": 40,
+    "B-postcode": 41,
+    "B-race_ethnicity": 42,
+    "B-religious_belief": 43,
+    "B-sexuality": 44,
+    "B-ssn": 45,
+    "B-state": 46,
+    "B-street_address": 47,
+    "B-swift_bic": 48,
+    "B-tax_id": 49,
+    "B-time": 50,
+    "B-unique_id": 51,
+    "B-url": 52,
+    "B-user_name": 53,
+    "B-vehicle_identifier": 54,
+    "I-PASSPORT_NUMBER": 109,
+    "I-PPSN": 107,
+    "I-account_number": 55,
+    "I-api_key": 56,
+    "I-bank_routing_number": 110,
+    "I-biometric_identifier": 57,
+    "I-blood_type": 58,
+    "I-certificate_license_number": 59,
+    "I-city": 60,
+    "I-company_name": 61,
+    "I-coordinate": 62,
+    "I-country": 63,
+    "I-county": 64,
+    "I-credit_debit_card": 65,
+    "I-customer_id": 66,
+    "I-date": 67,
+    "I-date_of_birth": 68,
+    "I-date_time": 69,
+    "I-device_identifier": 70,
+    "I-education_level": 71,
+    "I-email": 72,
+    "I-employee_id": 73,
+    "I-employment_status": 74,
+    "I-fax_number": 75,
+    "I-first_name": 76,
+    "I-gender": 77,
+    "I-health_plan_beneficiary_number": 78,
+    "I-http_cookie": 79,
+    "I-ipv4": 80,
+    "I-ipv6": 81,
+    "I-language": 82,
+    "I-last_name": 83,
+    "I-license_plate": 84,
+    "I-mac_address": 85,
+    "I-medical_record_number": 86,
+    "I-occupation": 87,
+    "I-password": 88,
+    "I-phone_number": 89,
+    "I-pin": 90,
+    "I-political_view": 91,
+    "I-postcode": 92,
+    "I-race_ethnicity": 93,
+    "I-religious_belief": 94,
+    "I-sexuality": 95,
+    "I-ssn": 96,
+    "I-state": 97,
+    "I-street_address": 98,
+    "I-swift_bic": 99,
+    "I-tax_id": 100,
+    "I-time": 101,
+    "I-unique_id": 102,
+    "I-url": 103,
+    "I-user_name": 104,
+    "I-vehicle_identifier": 105,
+    "O": 0
+  },
+  "max_position_embeddings": 512,
+  "model_type": "distilbert",
+  "n_heads": 12,
+  "n_layers": 6,
+  "num_span_labels": 11,
+  "output_past": true,
+  "pad_token_id": 0,
+  "qa_dropout": 0.1,
+  "seq_classif_dropout": 0.2,
+  "sinusoidal_pos_embds": false,
+  "span_label_max_span_tokens": {
+    "ACCOUNT_NUMBER": 19,
+    "BANK_ROUTING_NUMBER": 6,
+    "CREDIT_DEBIT_CARD": 13,
+    "EMAIL": 16,
+    "FIRST_NAME": 5,
+    "LAST_NAME": 8,
+    "PASSPORT_NUMBER": 9,
+    "PHONE_NUMBER": 10,
+    "POSTCODE": 8,
+    "PPSN": 9,
+    "SWIFT_BIC": 8
+  },
+  "span_label_names": [
+    "ACCOUNT_NUMBER",
+    "BANK_ROUTING_NUMBER",
+    "CREDIT_DEBIT_CARD",
+    "EMAIL",
+    "FIRST_NAME",
+    "LAST_NAME",
+    "PASSPORT_NUMBER",
+    "PHONE_NUMBER",
+    "POSTCODE",
+    "PPSN",
+    "SWIFT_BIC"
+  ],
+  "span_label_thresholds": {
+    "ACCOUNT_NUMBER": 0.5,
+    "BANK_ROUTING_NUMBER": 0.5,
+    "CREDIT_DEBIT_CARD": 0.5,
+    "EMAIL": 0.5,
+    "FIRST_NAME": 0.5,
+    "LAST_NAME": 0.5,
+    "PASSPORT_NUMBER": 0.5,
+    "PHONE_NUMBER": 0.5,
+    "POSTCODE": 0.5,
+    "PPSN": 0.5,
+    "SWIFT_BIC": 0.5
+  },
+  "span_positive_weight": 6.0,
+  "tie_weights_": true,
+  "token_extend_thresholds": {
+    "ACCOUNT_NUMBER": 0.08,
+    "BANK_ROUTING_NUMBER": 0.3,
+    "CREDIT_DEBIT_CARD": 0.3,
+    "EMAIL": 0.15,
+    "FIRST_NAME": 0.3,
+    "LAST_NAME": 0.3,
+    "PASSPORT_NUMBER": 0.15,
+    "PHONE_NUMBER": 0.15,
+    "POSTCODE": 0.3,
+    "PPSN": 0.3,
+    "SWIFT_BIC": 0.3
+  },
+  "token_focal_gamma": 1.0,
+  "token_hard_fraction": 0.4,
+  "token_label_thresholds": {
+    "ACCOUNT_NUMBER": 0.18,
+    "BANK_ROUTING_NUMBER": 0.8,
+    "CREDIT_DEBIT_CARD": 0.8,
+    "EMAIL": 0.95,
+    "FIRST_NAME": 0.3,
+    "LAST_NAME": 0.4,
+    "PASSPORT_NUMBER": 0.75,
+    "PHONE_NUMBER": 0.65,
+    "POSTCODE": 0.9,
+    "PPSN": 0.7,
+    "SWIFT_BIC": 0.8
+  },
+  "token_positive_weight": 4.0,
+  "token_presence_weight": 1.0,
+  "transformers_version": "4.57.6",
+  "vocab_size": 119547
+}

onnx/model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eaa4cd93ce1f81d78c46f09a7813c88f5bc3447f4694a0ff8326f452d36e6b8d
+size 539156462

onnx/model_quantized.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b62273cd056ccc4056e98d74133919492af0ddfc7acc29119ac7b7a06d50d6d6
+size 411983965

onnx/onnx_export.json

@@ -0,0 +1,12 @@
+{
+  "source_model": "models/irishcore-diffmask-135m-v1-rc6k-blend-0p30",
+  "onnx_path": "tmp_rc6k_onnx/model.onnx",
+  "task": "multitask-token-span-extraction",
+  "opset": 18,
+  "max_length": 256,
+  "output_names": [
+    "token_logits",
+    "start_logits",
+    "end_logits"
+  ]
+}

onnx/quantization.json

@@ -0,0 +1,26 @@
+{
+  "source_dir": "tmp_rc6k_onnx",
+  "input_model": "tmp_rc6k_q8/model.onnx",
+  "preprocessed_input_model": "",
+  "output_model": "tmp_rc6k_q8/model_quantized.onnx",
+  "weight_type": "qint8",
+  "per_channel": true,
+  "reduce_range": false,
+  "preprocess_applied": false,
+  "op_types": [
+    "MatMul",
+    "Gemm",
+    "Attention"
+  ],
+  "copied_assets": [
+    "tmp_rc6k_onnx/model.onnx",
+    "onnx_export.json",
+    "config.json",
+    "special_tokens_map.json",
+    "tokenizer.json",
+    "tokenizer_config.json",
+    "vocab.txt"
+  ],
+  "format": "onnx_dynamic_quantized",
+  "task": "token-classification"
+}

onnx/special_tokens_map.json

@@ -0,0 +1,37 @@
+{
+  "cls_token": {
+    "content": "[CLS]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "mask_token": {
+    "content": "[MASK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "[PAD]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "content": "[SEP]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "[UNK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

onnx/tokenizer_config.json

@@ -0,0 +1,64 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "100": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "101": {
+      "content": "[CLS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "102": {
+      "content": "[SEP]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "103": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": false,
+  "cls_token": "[CLS]",
+  "do_lower_case": false,
+  "extra_special_tokens": {},
+  "fix_mistral_regex": true,
+  "mask_token": "[MASK]",
+  "max_length": 512,
+  "model_max_length": 512,
+  "pad_to_multiple_of": null,
+  "pad_token": "[PAD]",
+  "pad_token_type_id": 0,
+  "padding_side": "right",
+  "sep_token": "[SEP]",
+  "stride": 0,
+  "strip_accents": null,
+  "tokenize_chinese_chars": true,
+  "tokenizer_class": "DistilBertTokenizer",
+  "truncation_side": "right",
+  "truncation_strategy": "longest_first",
+  "unk_token": "[UNK]"
+}

pyproject.toml

@@ -0,0 +1,17 @@
+[project]
+name = "irishcore-diffmask"
+version = "0.1.0"
+description = "IrishCore DiffMask token-span release with dynamic q8 ONNX export"
+requires-python = ">=3.10"
+readme = "README.md"
+license = { text = "Apache-2.0" }
+dependencies = [
+  "transformers>=4.41.0",
+  "torch",
+  "numpy>=1.26.0",
+  "onnxruntime>=1.20.0",
+  "huggingface_hub>=0.36.0",
+]
+
+[tool.uv]
+package = false

qa_config.json

@@ -0,0 +1,16 @@
+{
+  "release": "IrishCore-DiffMask-135M-v1-rc6",
+  "repo_id": "temsa/IrishCore-DiffMask-135M-v1-rc6",
+  "decoder": "raw_only_diffmask_token_span",
+  "diffusion_style_training": true,
+  "runtime_diffusion": false,
+  "min_score": 0.5,
+  "recommended_backend": "onnx_q8_cpu",
+  "onnx_file": "onnx/model_quantized.onnx",
+  "full_example": "My PPSN is 1234567TW, my Eircode is D02 X285, and my phone is 087 123 4567.",
+  "notes": [
+    "No scanner or validator layer is required for release behavior.",
+    "Use the bundled inference scripts or import common.decode_token_presence_segments.",
+    "The bundled scripts mask with [PII:LABEL] placeholders."
+  ]
+}

special_tokens_map.json

@@ -0,0 +1,37 @@
+{
+  "cls_token": {
+    "content": "[CLS]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "mask_token": {
+    "content": "[MASK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "[PAD]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "content": "[SEP]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "[UNK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer_config.json

@@ -0,0 +1,64 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "100": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "101": {
+      "content": "[CLS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "102": {
+      "content": "[SEP]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "103": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": false,
+  "cls_token": "[CLS]",
+  "do_lower_case": false,
+  "extra_special_tokens": {},
+  "fix_mistral_regex": true,
+  "mask_token": "[MASK]",
+  "max_length": 512,
+  "model_max_length": 512,
+  "pad_to_multiple_of": null,
+  "pad_token": "[PAD]",
+  "pad_token_type_id": 0,
+  "padding_side": "right",
+  "sep_token": "[SEP]",
+  "stride": 0,
+  "strip_accents": null,
+  "tokenize_chinese_chars": true,
+  "tokenizer_class": "DistilBertTokenizer",
+  "truncation_side": "right",
+  "truncation_strategy": "longest_first",
+  "unk_token": "[UNK]"
+}

training_sources.json

@@ -0,0 +1,86 @@
+{
+  "release": "IrishCore-DiffMask-135M-v1-rc6",
+  "base_model": "OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1",
+  "public_references": {
+    "rc5": "temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc5",
+    "rc8": "temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc8"
+  },
+  "task": "Irish core PII detection and masking in English and Irish Gaelic",
+  "coverage": [
+    "PPSN",
+    "ACCOUNT_NUMBER",
+    "BANK_ROUTING_NUMBER",
+    "CREDIT_DEBIT_CARD",
+    "PASSPORT_NUMBER",
+    "POSTCODE",
+    "PHONE_NUMBER",
+    "EMAIL",
+    "FIRST_NAME",
+    "LAST_NAME",
+    "SWIFT_BIC"
+  ],
+  "architecture": {
+    "family": "DistilBERT-size token-span extractor",
+    "diffusion_style_training": true,
+    "runtime_diffusion": false,
+    "scanner_free": true,
+    "validator_free": true,
+    "heads": [
+      "token_presence_head",
+      "typed_start_boundary_head",
+      "typed_end_boundary_head"
+    ]
+  },
+  "training_data": {
+    "published": [
+      "temsa/OpenMed-Irish-CorePII-TrainMix-v1",
+      "temsa/OpenMed-Irish-PPSN-Eircode-Spec-v1",
+      "joelniklaus/mapa",
+      "gretelai/synthetic_pii_finance_multilingual"
+    ],
+    "local_synthetic_hardening_sets": [
+      "irish_dllm_hardening_v1",
+      "dllm_gap_patch_v1",
+      "dllm_gap_patch_v2",
+      "dllm_gap_patch_v3",
+      "irish_core_diffmask_v2_mix",
+      "irish_core_diffmask_v3_mix",
+      "irish_core_diffmask_v4_mix"
+    ],
+    "selection_note": "The published checkpoint was selected from multiple continuation and interpolation runs to balance Irish core, multilingual PPSN, and hardening performance."
+  },
+  "training_recipe": {
+    "noise_schedule_family": "linear masked denoising schedule",
+    "runtime_diffusion": false,
+    "train_time_diffusion_steps": 4,
+    "start_noise_fraction": 0.65,
+    "end_noise_fraction": 0.05,
+    "loss": "average BCE losses over token presence and typed boundaries across noised passes"
+  },
+  "references": [
+    {
+      "title": "BERT",
+      "url": "https://arxiv.org/abs/1810.04805"
+    },
+    {
+      "title": "DistilBERT",
+      "url": "https://arxiv.org/abs/1910.01108"
+    },
+    {
+      "title": "Boundary Smoothing for Named Entity Recognition",
+      "url": "https://aclanthology.org/2022.acl-long.490/"
+    },
+    {
+      "title": "SPANNER: Named Entity Re-/Recognition as Span Prediction",
+      "url": "https://aclanthology.org/2021.acl-long.558/"
+    },
+    {
+      "title": "LLaDA 2.0: Scaling Up Diffusion Language Models to 100B",
+      "url": "https://arxiv.org/abs/2512.15745"
+    },
+    {
+      "title": "Scaling Diffusion Language Models via Adaptation from Autoregressive Models",
+      "url": "https://arxiv.org/abs/2410.17891"
+    }
+  ]
+}