You can deploy this model and start earning money today!

Ahahajij182u2/
local_doctor-360M

Name: local_doctor-360M
Availability: InStock

safetensors

llama

Views

No views yet

🩺 Local Doctor 360M – Model Card

📌 Model Details

Model Name: local_doctor-360M
Base Model: SmolLM2-360M-Instruct
Model Type: Causal Language Model (Instruction-tuned)
Framework: Transformers (PyTorch)
Fine-tuning Type: Domain-specific instruction tuning (medical symptom → condition mapping)
Developer: Independent / Local deployment focused

🧠 Model Description

Local Doctor 360M is a lightweight, locally deployable medical assistant model designed to:

Interpret user-provided symptoms
Suggest possible medical conditions
Run entirely on local machines (no cloud dependency)

The model is optimized for:

Privacy-first usage
Offline inference
Low-resource environments

🎯 Intended Use

✅ Primary Use Cases

Symptom-based preliminary health insights
Educational or experimental AI applications
Local/offline medical assistant prototypes
Research on small medical LLMs

❌ Not Intended For

Clinical diagnosis
Emergency decision-making
Replacement for licensed medical professionals

⚠️ Limitations & Risks

This model has significant limitations:

❗ May produce incorrect or misleading diagnoses
❗ Struggles with overlapping symptoms
❗ Limited reasoning due to small model size (360M parameters)
❗ May generate hallucinated or non-medical advice
❗ Not trained on real-time or verified clinical datasets

🚨 Safety Notice

This model must not be used for real medical decisions. Always consult a qualified healthcare professional.

📊 Performance Summary

Based on internal testing:

Approximate accuracy: ~40–50% on structured symptom datasets
Performs well on:
- Clear, high-signal symptom patterns (e.g., stroke, migraine)
Performs poorly on:
- Ambiguous cases
- Rare conditions
- Differential diagnosis scenarios

🧪 Example

Input

fever, headache, fatigue

Output

Possible conditions: Viral infection, Flu

🏗️ Training Details

Fine-tuned on:
- Symptom-condition mappings
- Instruction-style medical prompts
Training style:
- Supervised fine-tuning (SFT)
No reinforcement learning or alignment tuning applied

🔐 Privacy & Local Deployment

A key feature of this model:

✅ Runs 100% locally
✅ No internet required
✅ User data never leaves the device

⚙️ Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch


model_name = "Ahahajij182u2/local_doctor-360M"

# Load model and tokenizer once
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
model.eval()

def generate_diagnosis(symptoms: str, max_tokens: int = 300) -> str:
    """Generate diagnosis for given symptoms."""
    prompt = f"""<|im_start|>system
You are a medical assistant AI.

Rules:
- Predict the most likely disease from the symptoms.
- Give 3 to 5 basic precautions or first steps.
- Use only this format:

POSSIBLE DISEASE: ...

POSSIBLE PRECAUTIONS: ...

- If symptoms suggest an emergency, advise immediate medical help.
- Do not give unusual or unsafe advice.
<|im_end|>
<|im_start|>user
{symptoms}
<|im_end|>
<|im_start|>assistant
"""
    
    inputs = tokenizer(prompt, return_tensors="pt")
    
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_tokens,
            temperature=0.1,
            do_sample=True,
        )
    
    decoded = tokenizer.decode(
        outputs[0][inputs["input_ids"].shape[1]:],
        skip_special_tokens=False
    )
    
    return decoded.split("<|im_end|>")[0].strip()



# Test cases
test_symptoms = [
    "chest pain, shortness of breath",
    "diarrhea, vomiting, stomach pain",
    "rash, itching, swelling",
    "cough, fever, sore throat",
]

for symptoms in test_symptoms:
    response = generate_diagnosis(symptoms)
    print(f"\n{symptoms}\n→ {response}")
    assert "POSSIBLE DISEASE" in response
    assert "POSSIBLE PRECAUTIONS" in response

📦 Model Size & Requirements

Size: ~300–500 MB (depending on format)
Runs on:
- CPU ✅
- Low-end GPU ✅
Suitable for:
- Laptops
- Edge devices

🔧 Known Issues

Inconsistent formatting of outputs
May repeat prompt structure
Sometimes outputs unrelated conditions
Weak handling of multi-condition reasoning

🚀 Future Improvements

Better medical dataset curation
Instruction tuning for safer outputs
Multi-condition ranking
Confidence scoring
Integration with rule-based safety filters

📜 License

Based on the license of the base model: Refer to SmolLM2-360M-Instruct for details.

🙌 Acknowledgements

Built on top of Hugging Face ecosystem
Fine-tuned with Unsloth
Base model by HuggingFaceTB