hardrave/medicanite-granite-3.3-2b-instruct

Medicanite – Fine-Tuned Medical Language Model (Granite-3.3-2B-Instruct)

Model Summary

Medicanite is a fine-tuned, 2.53-billion parameter instruction-following model specialized for medical and healthcare-related tasks. Built on top of Granite-3.3-2B-Instruct, Medicanite is optimized for factuality and QA performance in clinical domains.

It was trained on the MedQUAD corpus (pre-cleaned) using 15,586 curated QA pairs for training and 821 for validation, with a low learning rate and LoRA (r=4, α=16) adaptation strategy. Despite its compact size, Medicanite outperforms both its base model and other leading open medical models in its class.

---

🔬 Benchmark Performance

Task	Medicanite	Granite-3.3-2B	JSL-MedPhi2-2.7B
MedMCQA	40.90%	40.50%	38.90%
MedQA	43.80%	42.50%	38.41%
MMLU Anatomy	53.30%	51.90%	48.89%
MMLU Clinical Knowledge	57.70%	60.80%	63.02%
MMLU College Biology	59.00%	59.00%	62.50%
MMLU College Medicine	53.80%	54.30%	58.38%
MMLU Medical Genetics	67.00%	62.00%	62.00%
MMLU Professional Medicine	54.00%	51.80%	44.12%
PubMedQA	72.80%	71.00%	73.00%
Average	55.82%	54.86%	54.36%

---

🛠️ Training Details

• Base model: ibm-granite/granite-3.3-2b-instruct
• Training corpus: MedQUAD (cleaned)
• # Training examples: 15,586
• # Validation examples: 821
• LoRA rank (r): 4
• Alpha: 16
• Learning rate: 7e-6
• Precision: bfloat16

---

🩺 Intended Use

This model is optimized for:

• Medical question answering
• Clinical knowledge tasks
• Instruction-following in health-related domains
• Educational use in biomedical fields
• Local/offline deployment for low-resource settings

---

⚠️ Limitations

• Not a diagnostic tool. Medicanite is intended for research and educational purposes only.
• May hallucinate facts or give outdated/incomplete medical advice.
• Always consult a licensed medical professional for clinical decisions.

---

📜 License

Apache 2.0 — same as the base Granite model.

---