| --- |
| language: |
| - de |
| - en |
| library_name: transformers |
| pipeline_tag: token-classification |
| base_model: bardsai/eu-pii-anonimization-multilang |
| tags: |
| - pii-detection |
| - named-entity-recognition |
| - privacy |
| - xlm-roberta |
| --- |
| |
| # Sano Shield 1.1 |
|
|
| Sano Shield 1.1 is Sano AI's multilingual privacy model for identifying sensitive |
| information before text reaches an AI provider. Its technical model ID is |
| `sano-shield-1-multilingual`. |
|
|
| > **Availability:** This repository publishes the model card only. Model weights, |
| > tokenizer files, ONNX artifacts, and training data are not publicly distributed or |
| > downloadable from Hugging Face. |
|
|
| ## Model details |
|
|
| | Field | Value | |
| |---|---| |
| | Model family | Sano Shield | |
| | Release | Sano Shield 1.1 | |
| | Technical ID | `sano-shield-1-multilingual` | |
| | Task | Token classification / sensitive-entity recognition | |
| | Evaluated languages | German and English | |
| | Entity classes | 35 entity types, 69 BIO token labels | |
| | Architecture | XLM-RoBERTa token classifier | |
| | Base-model lineage | [`bardsai/eu-pii-anonimization-multilang`](https://huggingface.co/bardsai/eu-pii-anonimization-multilang) | |
| | Release date | 2026-08-01 | |
|
|
| Sano Shield 1.1 continues a private Sano Shield 1.0 checkpoint. Sano Shield 1.0 was a |
| full-parameter continued fine-tune of the pinned base-model revision |
| `0e72e19f030ed4e661b1673e549af8e0dd176386`. Training targets come from Sano's |
| versioned synthetic corpus rather than predictions copied from a separate evaluator. |
|
|
| ## Intended use |
|
|
| Sano Shield 1.1 is designed as one component in a privacy pipeline: |
|
|
| 1. identify sensitive spans in German and English text; |
| 2. replace those spans with controlled placeholders before external model inference; and |
| 3. support re-identification only inside a trusted boundary. |
|
|
| Production deployments should combine the model with deterministic recognizers, policy |
| checks, monitoring, and fail-safe handling. Organization names remain user-selectable but |
| are not treated as mandatory natural-person PII by the release policy. |
|
|
| The model should not be used as the sole basis for legal-compliance decisions, medical |
| decisions, identity verification, employee monitoring, or irreversible actions affecting |
| a person. |
|
|
| ## Evaluation |
|
|
| The promoted checkpoint passed both synthetic checkpoint gates and a separately maintained |
| product-pipeline evaluation. The product evaluation data was held out from training. |
|
|
| | Product gate | Sano Shield 1.1 result | |
| |---|---:| |
| | German protected-span recall | 470 / 470 (100%) | |
| | English protected-span recall | 4,187 / 4,188 (99.976%) | |
| | Zero-touch documents | 29 / 30 | |
| | Raw false or unknown detections | 83, reduced from 141 | |
| | ONNX/PyTorch label agreement | 100% | |
|
|
| On the synthetic continuation evaluation, exact entity F1 improved by 0.0454 for German |
| and 0.0209 for English relative to the Sano Shield 1.0 parent, while privacy-span recall |
| was unchanged. |
|
|
| These results do not guarantee equivalent performance on other domains. The held-out |
| product evaluation measures Sano's release policy and complete product pipeline, including |
| deterministic rules; it should not be interpreted as a neural-model-only benchmark. |
|
|
| ## Training |
|
|
| - Parent: private Sano Shield 1.0 selected checkpoint |
| - Corpus: Sano Shield v0.4 synthetic continuation corpus |
| - Schedule: one epoch at a `1e-6` learning rate |
| - Composition: 62.5% replay, 25% positive contextual data, 4.2% contrastive data, and |
| 8.3% new negatives |
| - Loss: square-root inverse class weighting capped at `3.0`, with the outside label at |
| weight `1.0` |
|
|
| The continuation run restarted from Sano Shield 1.0 rather than from either rejected |
| candidate run. |
|
|
| ## Limitations |
|
|
| - German and English are the only release-policy target languages. |
| - Unseen document types, OCR errors, adversarial formatting, and ambiguous context can |
| produce missed or extra spans. |
| - No statistical detector can guarantee that all sensitive information is found. |
| - The model recognizes spans; it does not determine whether processing them is lawful or |
| satisfy a particular compliance regime. |
| - High-risk deployments require domain-specific, legally usable, human-reviewed evaluation. |
|
|
| ## Distribution |
|
|
| This public page is provided for transparency about the model's purpose, lineage, |
| evaluation, and limitations. It does not grant access to the checkpoint or corpus, and no |
| public redistribution license for those artifacts is granted here. Contact Sano AI for |
| deployment and access information. |
|
|
