Repository navigation
Releases
Releases · seotarek/medorqis-llm
MedOrqis-LLM v1.0.0 — Clinical Intelligence & Multi-Model Benchmark Release
MedOrqis-LLM v1.0.0 — Clinical Intelligence & Multi-Model Benchmark Release
MedOrqis-LLM is an open-weights, bilingual (Arabic & English) clinical foundation model fine-tuned on Applied-Innovation-Center/Karnak-6B-v1.0 (Qwen3 architecture) via QLoRA 4-bit NF4. It introduces a Dual-Track Neuro-Symbolic architecture designed to eliminate probabilistic arithmetic hallucinations in clinical risk calculations and drug dosing.
Key Highlights
- Dual-Track Architecture: Decouples narrative clinical reasoning in transformer weights from mathematical calculations executed via 12 certified deterministic engines.
- 0.00% Arithmetic Error Rate: Evaluated against non-linear exponent equations (CKD-EPI 2021, Bedside Schwartz, Cockcroft-Gault CrCl).
- 100.0% Pediatric Overdose Safety Cap: Strictly truncates weight-based pediatric dosages to adult single-dose limits.
- Multi-Dialect Arabic Triage: Fully normalizes Egyptian, Saudi/Gulf, Levantine, and Maghrebi acute symptom vernacular into standardized clinical entities (SNOMED-CT / ICD-10).
- Trade-to-Generic Pharmacology: Resolves regional trade names (Concor, Plavix, Cataflam, Doliprane) to active INN molecules and intercepts fatal drug-drug interactions.
- Sovereign On-Premise Footprint: 198 MB adapter running within 7.68 GB VRAM on a single commodity GPU (RTX 3060, Tesla T4, NVIDIA L4) with zero cloud data transmission.
Artifacts and Hub Links
- Hugging Face Model: https://huggingface.co/seotarek/medorqis-llm
- Hugging Face Dataset (50 Gold Scenarios): https://huggingface.co/datasets/seotarek/medorqis-clinical-instructions
- Interactive Comparison Portal: https://seotarek.github.io/medorqis-llm/
Verified Benchmark Comparison
| Evaluation Metric | MedOrqis-LLM (6B LoRA) | Karnak-6B (Raw Base) | BioMistral-7B | GPT-4o (Zero-Shot) |
|---|---|---|---|---|
| Arithmetic Calculation Error | 0.00% | 29.3% - 41.5% Drift | 24.8% Drift | 8.2% - 15.4% |
| Pediatric Overdose Ceiling | 100.0% Enforced | Failed (0.0%) | 35.0% Adherence | 78.0% Adherence |
| Tool Calling Syntax Validity | 100.0% Valid | Unsupported | Unsupported | 98.0% (Via API) |
| Arabic Dialect Comprehension | 98.0% | 58.0% | 18.0% | 86.0% |
| Local Air-Gapped Privacy | 100% On-Premise | 100% On-Premise | 100% On-Premise | Cloud API Only |
| Minimum Hardware Required | 7.68 GB VRAM | 14.2 GB VRAM | 15.8 GB VRAM | External Datacenter |
Fast Deployment (vLLM)
vllm serve Applied-Innovation-Center/Karnak-6B-v1.0 \
--enable-lora \
--lora-modules medorqis-llm=seotarek/medorqis-llm \
--max-model-len 4096 \
--gpu-memory-utilization 0.90 \
--port 8000