🦷 DentaInstruct-1.2B
DentaInstruct-1.2B is a fine-tuned instruction-following language model designed to assist with dental domain queries. It was trained using clinically focused questions and answers from the miriad/miriad-4.4M dataset, specifically filtered for the Dental & Oral Medicine specialty. The base model used is LiquidAI/LFM2-1.2B, optimized with the Unsloth fine-tuning library.
Model Details
- Architecture: LFM2-1.2B (decoder-only transformer)
- Base Model:
LiquidAI/LFM2-1.2B - Fine-tuning Library: Unsloth
- Instruction Format: Chat-style, formatted using
apply_chat_template - Trained On: Dental subset of
miriad-4.4M - Compute: Trained using Google Colab T4 (free tier)
Benchmark Summary
The model was evaluated on a curated set of dental prompts with rich terminology across endodontics, periodontics, prosthodontics, and oral surgery. Responses were assessed for fluency, accuracy, and domain relevance.
- ✅ Terminology Handling: Excellent coverage of dental-specific terms
- ✅ Instruction Following: Clear, context-aware responses
- ✅ Answer Structure: Consistently professional and clinically coherent
- ⚠️ Minor Hallucinations: A few outputs demonstrated factual drift in rare cases
Limitations & Warnings
This model was fine-tuned using the MIRIAD dataset, which comes with the following caution:
This model is trained on data that has not been manually reviewed by medical experts. It should not be used for diagnostic purposes or to inform medical decision-making. It is intended for research and educational purposes only.
- Not a substitute for professional dental care
- Do not use this model for clinical diagnosis or treatment advice
Intended Use
- Educational Q&A for dental students
- Conversational chatbots focused on oral health
- Research on clinical-domain instruction tuning
Citation
If you use this model or parts of it, please consider citing:
@misc{miriad2024,
title={MIRIAD: Medical Instructional Record with Interactions and Answers Dataset},
author={Xue, Yutong and others},
year={2024},
url={https://huggingface.co/datasets/miriad/miriad-4.4M}
}
Acknowledgements
- LiquidAI for the LFM2 model series
- Unsloth for training acceleration
- MIRIAD authors for the publicly available medical dataset