Purpose-built transformer models for compliance-critical entity extraction. Small footprint, high accuracy, runs entirely on-device.
A Domain-Specific Language Model (DSLM) is a language model pre-trained or fine-tuned on domain-specific corpora. DSLMs learn the vocabulary, terminology, and linguistic patterns unique to a particular field (healthcare, finance, legal, etc.), enabling higher accuracy on domain tasks than general-purpose models.
At Aare, we use DSLMs as the extraction layer in the Aare Edge verification pipeline. They convert unstructured text into structured facts that can be formally verified by Z3 Lite.
Pre-trained DSLMs for common compliance domains. Each model is optimized for on-device inference with minimal footprint.
Healthcare / Privacy
Detects all 18 HIPAA Safe Harbor PHI categories in clinical text. Handles voice transcriptions, OCR output, and free-form medical notes.
Financial Services
Extracts loan parameters, applicant data, and decision factors from underwriting text. Built for ECOA and fair lending compliance.
Privacy / GDPR / CCPA
Broad PII detection for privacy compliance across industries. Covers personal identifiers, financial data, and biometric markers.
Any Compliance Domain
Need entity extraction for a specialized domain? We train custom DSLMs on your data with your entity schema.
All Aare DSLMs are designed for edge deployment: small enough to run on mobile devices, fast enough for real-time inference, accurate enough for compliance-critical applications.
| Specification | HIPAA DSLM | Standard DSLMs |
|---|---|---|
| Architecture | DistilBERT (6-layer transformer) | DistilBERT or Phi-3-mini |
| Parameters | 67M | 67M - 3.8B |
| Model Format | CoreML (.mlpackage) | CoreML, ONNX, TensorFlow Lite |
| Quantization | FP16 | FP16 / INT8 / INT4 |
| Model Size | ~127 MB | 50 MB - 500 MB |
| Inference Time | <50ms | 20-100ms |
| Max Sequence Length | 512 tokens | 512-2048 tokens |
| Tokenizer | WordPiece (BERT vocab) | WordPiece or BPE |
| Training Method | Fine-tuned NER with BIO tagging | Fine-tuned NER / LoRA |
A 67M parameter DSLM fits in ~127MB. GPT-4 class models require 100GB+. DSLMs run on phones, watches, and embedded devices.
Sub-50ms inference vs. seconds for cloud LLM calls. Real-time entity extraction without network latency.
Fine-tuned on domain-specific data, DSLMs outperform general models on their target entity types.
No temperature, no sampling variation. Same input always produces same entities. Critical for compliance.
All inference runs locally. PHI, PII, and sensitive data never transmitted. True on-device privacy.
No network required. Extract entities in airgapped environments, airplanes, or areas with poor connectivity.
Aare DSLMs are available under commercial license. Model weights are proprietary and optimized for specific compliance domains. The Aare Edge SDK (inference runtime, Z3 Lite, tokenizer) is MIT licensed and open source.
SDK is MIT licensed. Model weights under commercial license.
GitHub: https://github.com/aare-ai