Llama-3.2-1B-Instruct-Medicina-Generale/README.md

license, language, tags, datasets, base_model

license	language	tags	datasets	base_model
cc-by-nc-4.0	it	llama llama-3 meta medical-qa italian biomedical question-answering fine-tuning unsloth bnb 4bit imb Medicina-Generale	praiselab-picuslab/IMB	unsloth/Llama-3.2-1B-Instruct-unsloth-bnb-4bit

🧠 Llama-3.2-1B-Instruct — IMB Medicina Generale Fine-Tuned Model

This model is a fine-tuned version of unsloth/Llama-3.2-1B-Instruct-unsloth-bnb-4bit, optimized for Italian medical question answering, with a specific focus on Medicina Generale.

The fine-tuning was performed using a subset of the IMB (Italian Medical Benchmark) dataset, specifically:

• Medicina Generale category only
• ~10,000 training samples

The training was performed using the Unsloth library with LoRA fine-tuning, and the adapter weights were later merged into the base model to provide a standalone checkpoint.

This model relies on data from the IMB dataset. If you use this model in research or applications, you must cite the IMB paper (see Citation section below).

---

📚 Training Dataset — IMB (Italian Medical Benchmark)

IMB is an Italian benchmark for medical question answering, designed to evaluate and improve LLM performance in clinical-domain Italian language understanding and reasoning.

The full dataset includes:

• IMB-QA: 782,644 doctor-patient conversations collected from Italian online medical forums
• IMB-MCQA: 25,862 multiple-choice questions derived from Italian medical specialization exams

⚠️ Important: This model was trained only on the Medicina Generale subset (~10,000 samples) of IMB, not on the full dataset.

Dataset repository: 👉 https://github.com/PRAISELab-PicusLab/IMB

---

🧪 Usage Example

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("praiselab-picuslab/Llama-3.2-1B-Instruct-Medicina Generale")
tokenizer = AutoTokenizer.from_pretrained("praiselab-picuslab/Llama-3.2-1B-Instruct-Medicina Generale")

prompt = "[Example question in Italian about Medicina Generale]"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=150)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

⚠️ Usage Restrictions

• Allowed use: Non-commercial research only
• Redistribution: Not allowed without explicit authorization
• Mandatory citation: The IMB dataset paper must be cited in any publication or derived work

---

📄 Citation

If you use this model, the IMB dataset, or derived outputs in research, please cite:

@inproceedings{DBLP:conf/clic-it/RomanoRBPM25,
  author       = {Antonio Romano and
                  Giuseppe Riccio and
                  Mariano Barone and
                  Marco Postiglione and
                  Vincenzo Moscato},
  editor       = {Cristina Bosco and
                  Elisabetta Jezek and
                  Marco Polignano and
                  Manuela Sanguinetti},
  title        = {{IMB:} An Italian Medical Benchmark for Question Answering},
  booktitle    = {Proceedings of the Eleventh Italian Conference on Computational Linguistics
                  (CLiC-it 2025), Cagliari, Italy, September 24-26, 2025},
  series       = {{CEUR} Workshop Proceedings},
  volume       = {4112},
  publisher    = {CEUR-WS.org},
  year         = {2025},
  url          = {https://ceur-ws.org/Vol-4112/92_main_long.pdf}
}

---

🏗 Training Details

• Base model: unsloth/Llama-3.2-1B-Instruct-unsloth-bnb-4bit
• Fine-tuning method: LoRA (Unsloth)
• Quantization: 4-bit (BitsAndBytes)
• Adapter merging: Yes (Full merged model)
• Language: Italian
• Domain: Medical — Medicina Generale
• Training size: ~10,000 samples

---

📜 License

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License.

---

🤝 Acknowledgements

👨‍💻 This project was developed by Mariano Barone, Roberta Di Marino, Francesco Di Serio, Giovanni Dioguardi, Marco Postiglione, Antonio Romano, Giuseppe Riccio, and Vincenzo Moscato at University of Naples, Federico II