A medical conversation assistant LoRA fine-tuned to provide safe, calm, and structured responses to patient health questions, while avoiding diagnosis, prescriptions, and high-risk medical claims.

This version focuses on patient-style interactions, improving upon general medical Q&A by handling conversational cues, clarifying questions, and safety-oriented guidance.

It responds to medical questions with:

Calm, structured explanations
Encouragement for professional evaluation
High-level guidance on general health
Clarifying questions to better understand the patient
Avoids guessing diagnoses or giving medications

🧠 Base Model

Base: mistralai/Mistral-7B-Instruct-v0.2
Finetuning: LoRA (QLoRA low-rank adapter)
Framework: Unsloth + TRL + PEFT

This model does not perform diagnosis and does not provide medication dosing.

✅ Intended Use

Use Case	Allowed	Notes
General health questions	✅	Calm, structured guidance
Explaining symptoms conceptually	✅	High-level guidance, no diagnosis
First-aid general instructions	✅	Only general safety advice
Patient support conversation	✅	Reassuring and clarifying
Medical education support	✅	Suitable for study / learning
Diagnosing medical conditions	❌	Must avoid guessing
Prescribing medications	❌	High-risk → blocked behavior
Emergency care instructions	❌	Always refer to real professionals

⚖️ Safety Behavior

The model is trained to:

Avoid firm diagnoses
Encourage professional evaluation when appropriate
Ask clarifying questions
Not provide medication dosages
Avoid urgent-care decision instructions

This aligns the model with safe conversational medical use.

🏋 Training Details

Dataset Used

Dataset	Size	Purpose
`mahfoos/Patient-Doctor-Conversation` (filtered)	~3k examples	Base conversational patterns
Custom Gold Conversations	~80 examples	Style & safe tone fine-tuning

Data Preparation

Removed incomplete doctor responses
Reformatted dataset into prompt–response style
Emphasized safe communication guidelines

Training Configuration

Method: LoRA Finetuning
Rank: 16
Precision: 4-bit QLoRA
Batching: Gradient accumulation enabled for low-VRAM training

🧪 Example Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "mistralai/Mistral-7B-Instruct-v0.2"
adapter = "ysn-ir/mistral-medical-chat-lora-v2"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

prompt = "Patient: I have chest tightness and shortness of breath.\nDoctor:"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

🔔 Limitations

Not a medical professional
Does not know patient history
Risk of hallucination is mitigated, not removed
Should never be used for emergency treatment

If the user reports:

Severe pain
Difficulty breathing
Loss of consciousness
Chest pain

→ The model must advise immediate medical attention.

👤 Author

Hugging Face: https://huggingface.co/ysn-ir Maintained by: ysn-ir

🧾 License

Released under Apache-2.0. Permitted for:

Research
Personal use
Commercial use

Not permitted for clinical decision-making.

✨ Summary

This v2 model:

Focuses on patient-style conversation with safe guidance
Provides structured, calm responses for health-related queries
Reinforces professional referral behavior

v1 builds on this with education-focused Q&A rather than patient conversation.

Use responsibly.