abirmed_core_slm — Foundational General Medical Transformer

Part of the A.B.I.R Ecosystem

abirmed_core_slm is a foundational medical language model developed as part of the A.B.I.R Ecosystem and the ABIRMED Modular Medical Specialist Transformer System, a distributed artificial intelligence architecture designed to replicate real-world medical specialization using multiple specialist transformer models.

This model serves as the General Physician (GP) brain of the ABIRMED ecosystem, providing broad medical knowledge, clinical reasoning foundations, and general medical understanding that supports specialized medical transformer models.

This is Version 1, with future versions planned for expanded intelligence, higher accuracy, and deeper medical specialization.

ABIRMED — Modular Medical Specialist Transformer System

ABIRMED is a modular medical AI ecosystem consisting of multiple specialist Small Language Models (SLMs), each trained for a specific medical domain. Instead of using a single large monolithic model, ABIRMED uses a distributed specialist architecture inspired by real-world clinical specialization.

Each model acts as an independent medical specialist while collectively forming a unified medical reasoning system.

This modular approach provides:

Higher accuracy within specialized domains
Lower computational requirements
CPU-efficient inference capability
Scalable and extensible medical intelligence architecture

Developed by: Abir Maheshwari
Architecture: Modular Decoder-only Transformer System
Framework: PyTorch + HuggingFace Transformers
Training Platform: Google Colab T4 GPU
License: MIT

Role of abirmed_core_slm in the ABIRMED System

abirmed_core_slm functions as the General Medical Intelligence Layer, equivalent to a General Physician in real-world healthcare.

It provides:

General medical knowledge
Disease explanation capability
Foundational clinical reasoning
Medical education support
Core reasoning support for specialist ABIRMED models

This model acts as the foundation upon which specialist models such as diagnosis, pharmacology, pathology, emergency, psychiatry, dermatology, cardiology, pediatrics, and veterinary models operate.

Model Details

Model Name: abirmed_core_slm
Version: 1.0
Developer: Abir Maheshwari
Organization: A.B.I.R Ecosystem
Model Type: Causal Language Model (Decoder-only Transformer)
Architecture: Custom Transformer built from scratch
Base Model: None
License: MIT

Technical Specifications

Architecture: Decoder-only Transformer

Parameters: ~38 Million

Transformer Layers: 8

Attention Heads: 8

Hidden Size: 512

Intermediate Size: 2048

Context Length: 256 tokens

Tokenizer: GPT-2 tokenizer with custom PAD token

Weight Sharing: Embedding and LM Head tied

Training Objective: Causal Language Modeling

Precision: FP16 mixed precision

Framework: PyTorch

Export Formats:

safetensors
PyTorch (.pt)

Checkpoint Support:

Full training state resume support

Training Details

Training Dataset

Primary dataset used:

lavita/MedQuAD

This dataset contains structured medical question-answer pairs covering:

Diseases
Symptoms
Treatments
Medical definitions
Clinical explanations

This dataset was selected to provide strong foundational medical knowledge.

Training Procedure

Optimizer: AdamW

Learning Rate: 5e-4

Batch Size: 8

Gradient Accumulation Steps: 2

Training Platform:

Google Colab
NVIDIA T4 GPU

Training Objective:

Predict next token in medical text sequences

Training Duration:

Optimized for modular incremental training cycles

Capabilities

abirmed_core_slm is capable of:

Explaining diseases
Answering general medical questions
Providing medical education explanations
Understanding general clinical concepts
Supporting medical reasoning systems

Example:

Input: "What is diabetes?"

Output: "Diabetes is a chronic medical condition where the body cannot properly regulate blood sugar levels due to insulin dysfunction."

Intended Use

This model is intended for:

Medical AI research
Educational medical assistants
Medical chatbot development
AI healthcare research
Offline medical AI systems

Out-of-Scope Use

This model is not intended for:

Clinical diagnosis
Medical treatment decisions
Emergency medical care
Prescription decisions

This is a research model only.

Limitations

abirmed_core_slm:

Is not a licensed medical system
May produce incorrect medical information
May lack specialist-level reasoning
Should not replace healthcare professionals

Design Philosophy

ABIRMED follows a modular specialist architecture rather than a monolithic model.

This approach allows:

Higher domain accuracy
Lower hallucination rates
Better reasoning specialization
Efficient CPU deployment

Each model acts as a specialist doctor.

abirmed_core_slm acts as the General Physician.

A.B.I.R Ecosystem Integration

abirmed_core_slm is part of the broader A.B.I.R Ecosystem, which includes:

Language models
Medical AI systems
Domain-specific transformer models
Modular AI intelligence systems

ABIRMED represents the medical intelligence division of the A.B.I.R Ecosystem.

Version

Version: 1.0

Future versions will include:

Improved reasoning accuracy
Expanded datasets
Longer context length
Enhanced clinical intelligence

Author

Abir Maheshwari
Independent AI Researcher
Founder, A.B.I.R Ecosystem

Hugging Face:
https://huggingface.co/abirmaheshwari

License

MIT License

Downloads last month: 14