---
base_model:
- mistralai/Mistral-7B-Instruct-v0.3
library_name: peft
tags:
- medical
- instruction-tuning
- dental
- lora
- mistral
- unsloth
license: mit
language:
- en
---

# 🦷 LoRA Fine-Tuned Model for Dental AI Assistant (Stage 1 - Tone Fine-Tuning)

This is a LoRA adapter trained on conversational medical data to teach a base language model (Mistral-7B-Instruct) how to behave like a helpful and professional medical assistant, focusing on polite tone, empathetic responses, and clarity.

This model is intended for **Stage 1 tone/behavior fine-tuning** in a two-phase pipeline toward building an AI dental doctor.

---

## 📋 Model Details

- **Developed by:** BirdieByte
- **Finetuned from:** `unsloth/mistral-7b-instruct-v0.3-bnb-4bit`
- **LoRA Framework:** [PEFT](https://github.com/huggingface/peft)
- **Training Library:** [Unsloth](https://github.com/unslothai/unsloth)
- **Language(s):** English
- **Model Type:** LoRA adapter for causal language modeling
- **License:** MIT

---

## 💡 Intended Use

### Direct Use

This model adapter is used to:
- Enhance tone and medical empathy in LLMs
- Simulate a virtual assistant that responds like a professional doctor
- Prepare a model for further domain-specific fine-tuning (e.g. dental implants)

### Not for Use

- Not intended for medical decision-making or diagnosis.
- Not safe to use in real-world healthcare settings without human supervision.

---

## 🧠 Training Details

- **Training Data:** Reformatted and filtered samples from [`antareepdey/Patient_doctor_chat`](https://huggingface.co/datasets/antareepdey/Patient_doctor_chat)
- **Data Format:** ChatML-style (ChatGPT-like role-based prompting)
- **Training Type:** LoRA adapter tuning only (base model not updated)
- **Precision:** 4-bit base model (bnb), LoRA adapter in FP16
- **Steps:** 3 epochs on ~480 examples
- **Output:** LoRA adapter weights and tokenizer

---

## 🧪 Evaluation

Qualitative testing showed:
- Improved clarity
- Friendly tone and structure
- Useful for staged evaluation against baseline Mistral model

Quantitative metrics not reported at this stage.

---

## 🚀 How to Use

### Load with PEFT

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.1")
model = PeftModel.from_pretrained(base, "BirdieByte1024/Mistral-doctor-behavior")
tokenizer = AutoTokenizer.from_pretrained("BirdieByte1024/Mistral-doctor-behavior")
```

---

## 🌱 Environmental Impact

- **Hardware:** Google Colab (T4 GPU)
- **Training time:** ~30 minutes
- **Energy usage:** Estimated minimal for LoRA
- **CO2 emissions:** < 0.01 kg CO2eq (estimated)

---

## 📚 Citation

BirdieByte (2025). *LoRA Fine-Tuned Dental Tone Assistant (Stage 1)*. Hugging Face model repository.

---

## ✉️ Contact

For questions or collaborations, please contact via Hugging Face (@your-username).

- PEFT 0.14.0