![[360Labs.ai]](/logo-new.svg){kind=logo}
{ sanad }· Released
Sanad-1.0
Clinical AI Assistant for Medical Diagnosis, Transcription, and Arabic Medical Support
// overview
Sanad-1.0 is a fine-tuned clinical AI model built on Google's MedGemma-27B-text-it. Purpose-built for Mediscribe, it provides clinical transcription (dialogue to SOAP notes), primary diagnosis extraction with ICD-10 coding, and ranked differential diagnosis generation. Trained on 551,491 curated medical examples across 15 specialized datasets using a 4-stage progressive QLoRA pipeline.
View on Hugging Face →// specs
Specifications
27B
Parameters
MedGemma-27B
Base Model
551K
Training Examples
87.7%
MedQA (USMLE)
90%
EHRQA Accuracy
128K tokens
Context Window
English + Arabic
Languages
BF16
Precision
32
LoRA Rank
21+
Specialties
QLoRA (4-bit NF4)
Fine-Tuning
8,192 tokens
Output Limit
// architecture
Architecture
1 Stage 1: Transcription (152,930 examples) - Dialogue to SOAP notes, clinical documentation2 Stage 2: Diagnosis (61,920 examples) - Clinical knowledge, diagnostic reasoning, ICD-103 Stage 3: Differential Diagnosis (129,696 examples) - Ranked DDx with chain-of-thought reasoning4 Stage 4: Arabic + Reinforcement (179,373 examples) - Bilingual support, existing clinical data
// features
Features
- 01Clinical transcription: converts doctor-patient conversations into structured SOAP notes and discharge summaries
- 02Primary diagnosis extraction with ICD-10 classification and clinical reasoning
- 03Ranked differential diagnosis with probability assessments and reasoning chains
- 04Chain-of-thought reasoning with transparent <thinking> traces
- 05Full bilingual capability (English + Arabic) for clinical consultations
- 06USMLE-level knowledge across 21+ medical specialties
- 07Structured JSON output for EHR integration
- 08128K context window for complete clinical transcripts and multi-visit patient histories
- 094-stage progressive fine-tuning with decreasing learning rates to prevent catastrophic forgetting
- 1087.7% MedQA accuracy, surpassing models with 2.5x more parameters
// benchmarks
Benchmarks
| Dataset | Score | Comparison |
|---|---|---|
| MedQA (USMLE) | 87.7% | Base Gemma 3 27B: 74.9% |
| MedMCQA | 74.2% | Base: 62.6% |
| MMLU-Med | 87.0% | Base: 83.3% |
| EHRQA (v1.5) | 90.0% | Base: 68.0% |
| AgentClinic-MedQA | 56.2% | Human physicians: 54% |
| AfriMed-QA | 84.0% | Base: 72.0% |
// end of modelSanad-1.0