metadata
license: apache-2.0
language:
- ar
- en
library_name: transformers
Brain Map AI v3.0 โ Clinical Intelligence Model
Brain Map AI is a specialized medical reasoning model optimized for clinical decision support, USMLE-style reasoning, and bilingual (English/Arabic) medical tasks.
๐ Deployment Status
The model is officially live on the gijl/ai repository. It features a technical stack including PyTorch, Hugging Face Transformers, and a specialized GGUF compatibility layer.
Key Features
- Architecture: Specialized Transformer with Grouped Query Attention (GQA) and Sparse Mixture of Experts (MoE) thinking blocks.
- Memory: Adaptive clinical memory using secure JSON/Safetensors storage (migrated from pickle for maximum security).
- Safety: Built-in Sovereign Critic and Drug Dosage Validation layers.
- Compatibility: Cross-compatible with Hugging Face Transformers and GGUF (via llama.cpp/LM Studio).
Technical Specifications
- Core: PyTorch 2.x / Transformers 5.x
- Precision: F16 (4-bit/8-bit quantization recommended for consumer VRAM optimization).
- Tokenizer: Custom BPE (8k-32k vocab) supporting medical nomenclature in EN/AR.
Usage
1. Using the Universal Loader (Recommended)
We provide a loader.py script for seamless integration across formats:
from loader import load_brain_map
model = load_brain_map('gijl/ai')
2. Standard Transformers Loading
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('gijl/ai')
model = AutoModelForCausalLM.from_pretrained('gijl/ai', trust_remote_code=True)
Unresolved Tasks & Next Steps
- VRAM Optimization: Further quantization (GGUF Q4_K_M) for mobile deployments.
- Extended Evaluation: Continued benchmarking against MedHALT and USMLE Step 2 vignettes.
Disclaimer: This model is for clinical research and decision-support assistance only. It is not a replacement for professional medical advice.