Lbai-1-preview is a specialized medical language model designed to assist with medical and healthcare-related queries, providing informative responses based on medical knowledge.

This model has been trained using advanced machine learning techniques to create a standalone, fully-integrated language model optimized for medical domain applications.

Developed by: Neurazum
Model type: Causal Language Model (Medical Domain)
Architecture: Transformer-based decoder model
Language: English
License: MIT
Model size: ~5GB (float16)
Parameters: 2.7B parameters

Intended Use

Direct Use

Lbai-1-preview is designed for:

Medical Question Answering: Providing informative responses to general medical questions
Medical Information Retrieval: Helping users understand medical concepts, symptoms, and conditions
Healthcare Education: Supporting medical students and healthcare professionals with educational content
Patient Education: Offering accessible explanations of medical terms and conditions
Clinical Documentation Assistance: Helping draft medical notes and summaries (with human oversight)
Medical Research Support: Assisting in literature review and medical knowledge synthesis

Recommended Use Cases

Educational platforms for medical training
Healthcare chatbots with appropriate disclaimers
Medical information systems with human-in-the-loop verification
Research assistants for medical literature analysis
Clinical documentation support tools

Use with Supervision

IMPORTANT: This model should ALWAYS be used with appropriate human supervision, especially in:

Clinical decision-making contexts
Patient care situations
Medication recommendations
Diagnosis suggestions
Treatment planning

Out-of-Scope Uses

Prohibited Uses

This model should NOT be used for:

Primary Clinical Diagnosis: Making medical diagnoses without professional medical oversight
Treatment Decisions: Making independent treatment or medication decisions
Emergency Medical Situations: Providing guidance in life-threatening or emergency situations
Replacing Healthcare Professionals: Substituting for qualified medical practitioners
Self-Medication Guidance: Providing medication recommendations without medical supervision
Critical Care Decisions: Making decisions in intensive care or critical medical situations
Legal Medical Advice: Providing legally binding medical opinions
Prescription Decisions: Determining medication dosages or prescriptions

High-Risk Scenarios

Pediatric Care: Specialized medical care for children requires expert oversight
Oncology: Cancer diagnosis and treatment planning
Surgery Planning: Surgical procedures and planning
Mental Health Crisis: Acute mental health emergencies
Rare Diseases: Conditions requiring specialized expertise

Ethical Boundaries

Generating fake medical records
Impersonating healthcare professionals
Providing medical advice in jurisdictions where AI cannot legally do so
Using for insurance fraud or medical billing manipulation

Training Data

Dataset Information

Training Dataset Composition:

Primary Source(s): Medical textbooks, PubMed articles, NCBI
Dataset Size: 4.8M token
Language(s): English and Turkish
Domain Coverage: Neuroscience, Neurology

Example Dataset Types:

Medical literature and textbooks
Clinical guidelines and protocols
Patient-doctor conversation datasets
Medical question-answer pairs
Clinical case studies
Medical terminology databases

Evaluation Data

Test Dataset

Evaluation Dataset:

Source: PubMed articles, NCBI
Size: 500K token
Split Ratio: 80% train, 10% validation, 10% test

Benchmark Results

Medical Knowledge Tests

Benchmark	Score
MedQA	36.53% (± 1.35%)
MedMCQA	35.16% (± 0.07%)
MMLU-Medical (Anatomy)	48.88% (± 0.43%)
MMLU-Medical (Clinical Knowledge)	59.62% (± 0.30%)
MMLU-Medical (College Biology)	58.33% (± 0.41%)
MMLU-Medical (College Medicine)	54.91% (± 0.37%)
MMLU-Medical (Medical Genetics)	61.00% (± 0.49%)
MMLU-Medical (Professional Medicine)	50.00% (± 0.30%)

General Language Tests

Benchmark	Score
MMLU (Overall)	54.26% (± 0.40%)
TruthfulQA	44.48% (± 1.52%)
HellaSwag	57.19% (± 0.49%)

Qualitative Analysis

Strengths:

Strong performance on general medical questions
Good understanding of medical terminology

Weaknesses:

May struggle with rare diseases
Limited multilingual support

Ethical Considerations

Medical AI Ethics

Patient Safety First: This model is designed as an assistive tool, not a replacement for healthcare professionals. All outputs should be:

Verified by qualified medical professionals before clinical use
Used only in appropriate contexts with proper oversight
Accompanied by clear disclaimers about AI limitations

Bias and Fairness

Potential Biases:

Dataset Bias: Medical literature may overrepresent certain populations or conditions
Language Bias: Model may perform differently across languages or dialects
Demographic Bias: Performance may vary across age groups, genders, or ethnicities
Geographic Bias: Medical practices and terminology vary by region

Mitigation Efforts:

Diverse evaluation datasets
Regular bias audits
Transparent reporting of limitations

Privacy and Data Protection

Training Data Privacy:

All training data has been processed to remove personally identifiable information (PII)
No patient health information (PHI) from the training data can be extracted from the model

User Privacy:

Users should not input sensitive personal health information
Conversations should be treated as non-confidential unless properly secured
Organizations deploying this model must implement appropriate privacy safeguards

Transparency and Accountability

Model Limitations:

This model can make mistakes and generate incorrect information
Medical knowledge evolves; model may contain outdated information
Model cannot access real-time medical databases or patient records
Cannot perform physical examinations or diagnostic tests

Accountability:

Responsibility for medical decisions remains with healthcare professionals
Users should verify all information with qualified sources
Feedback and error reporting mechanisms should be established

Societal Impact

Positive Impacts:

Improved access to medical information
Support for medical education
Assistance in underserved regions (with appropriate oversight)
Reduced burden on healthcare systems for routine inquiries

Potential Risks:

Over-reliance on AI for medical decisions
Spread of medical misinformation if used improperly
Reduced human interaction in healthcare
Unequal access based on digital divide

Recommendations:

Always use with professional medical oversight
Implement strong disclaimer systems
Provide user education on AI limitations
Ensure equitable access to both AI tools and human healthcare

Limitations

Technical Limitations

Knowledge Cutoff:
- Training data has a cutoff date Oct-2025
- May not include latest medical research and guidelines
- Drug information and treatment protocols may be outdated
Context Window:
- Maximum context length: 2048 tokens
- Cannot process very long medical documents in a single pass
- May lose context in extended conversations
Language Support:
- May have reduced performance in non-primary languages
- Medical terminology translation may be imperfect
Computational Requirements:
- Requires ~5GB VRAM for inference (float16)
- May be slow on CPU-only systems
- Real-time applications may need optimization

Medical Knowledge Limitations

Scope of Knowledge:
- Cannot access patient-specific data or medical records
- No real-time drug interaction databases
- Cannot perform calculations requiring patient-specific parameters
Diagnostic Limitations:
- Cannot perform physical examinations
- Cannot interpret medical images, lab results, or vital signs
- Diagnosis requires comprehensive clinical assessment beyond text
Specialization:
- May perform better in some medical specialties than others
- Rare diseases and conditions may not be well-represented
- Cutting-edge treatments may not be included
Clinical Reasoning:
- Cannot replace clinical judgment and experience
- May not consider all relevant factors in complex cases
- Cannot account for individual patient circumstances

Accuracy and Reliability

Hallucination Risk:
- Model may generate plausible-sounding but incorrect information
- Medical facts should always be verified
- Citations may be fabricated
Inconsistency:
- May provide different answers to similar questions
- Response quality may vary based on phrasing
- Temperature settings affect consistency
Uncertainty Calibration:
- Model may not accurately express uncertainty
- Confident-sounding responses may be incorrect
- No built-in confidence scores

Safety Limitations

Emergency Situations:
- ABSOLUTELY NOT for medical emergencies
- Cannot call emergency services
- Cannot provide real-time critical care guidance
Medication Safety:
- Cannot verify drug interactions for specific patients
- Cannot determine appropriate dosages
- Cannot account for allergies or contraindications
Mental Health:
- Not designed for crisis intervention
- Cannot provide ongoing mental health therapy
- Should not replace mental health professionals

Regulatory and Legal Limitations

Not FDA Approved:
- This model is not approved as a medical device
- Not certified for clinical decision support
- Not validated for regulatory compliance
Professional Liability:
- Does not establish doctor-patient relationship
- Users responsible for verifying all information
- Not a substitute for professional medical advice
Geographic Limitations:
- Medical practices vary by country and region
- May not align with local medical standards
- Regulatory status varies by jurisdiction

Mitigation Strategies

For Developers:

Implement hallucination detection systems
Add fact-checking layers
Provide confidence scores when possible
Regular model updates with new medical knowledge

For Users:

Always verify critical information
Use as a supplementary tool, not primary source
Consult qualified healthcare professionals
Report errors and inconsistencies

For Organizations:

Implement human-in-the-loop systems
Regular audits and quality checks
Clear disclaimers and user warnings
Appropriate training for staff using the model

How to Use

Basic Usage

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model_path = "Neurazum/Lbai-1-preview"  # Update with your actual path
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)

tokenizer = AutoTokenizer.from_pretrained(model_path)

# Example usage
prompt = "Patient: I have a persistent headache and fever. What could be the cause?\nDoctor:"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=200,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

System Requirements

Minimum:

GPU: 6GB VRAM (float16)
RAM: 8GB
Storage: 6GB

Recommended:

GPU: 8GB+ VRAM
RAM: 16GB+
Storage: 10GB

Installation

pip install torch transformers accelerate

Medical Disclaimer

⚠️ IMPORTANT MEDICAL DISCLAIMER ⚠️

This AI model is provided for informational and educational purposes only and is NOT a substitute for professional medical advice, diagnosis, or treatment.

Key Points:

Not Medical Advice: The information provided by this model does not constitute medical advice and should not be relied upon for medical decisions.
Consult Healthcare Professionals: Always seek the advice of your physician or other qualified health provider with any questions you may have regarding a medical condition.
Emergency Situations: Never disregard professional medical advice or delay in seeking it because of something you have read or received from this AI model. If you think you may have a medical emergency, call your doctor or emergency services immediately.
No Doctor-Patient Relationship: Use of this model does not create a doctor-patient relationship between you and the developers or operators of this model.
Accuracy Not Guaranteed: While efforts have been made to ensure accuracy, medical information can change rapidly, and this model may contain errors or outdated information.
Individual Variation: Medical conditions and treatments vary greatly among individuals. Information that may be appropriate for one person may not be appropriate for another.
Verify Information: Always verify any medical information with qualified healthcare professionals and current medical literature.

By using this model, you acknowledge and agree to these limitations and disclaimers.

Citation

If you use this model in your research or applications, please cite:

bibtex

@misc{neurazum,
   title = {Lbai-1-preview},
   author = {Neurazum AI Department},
   year = {2025},
   url = {https://huggingface.co/Neurazum/Lbai-1-preview},
}

Contact

Organization: Neurazum
Contact Email: contact@neurazum.com
Website: neurazum.com

Updates and Versions

Version 1.0-preview (Current)

Initial release
Medical domain optimization

Planned Updates

[×] Additional medical domain training
[×] Multilingual support expansion
[×] Safety improvements
[×] Integration of bai-Mind models that interpret EEG signals
[×] Reasoning ability

Last Updated: 2025-11-25 Model Version: 1.0-preview Documentation Version: 1.0