mistral-medical-chat-lora-v2

🩺 Mistral-Medical-Chat-LoRA v2

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

---

⚖️ 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

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

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.