The challenge 🎯
Understand how data quality, model scaling, and inference optimization directly determine model behavior for a specific linguistic context.
← Back to projects
Project spotlight ✨
Stambecco
Stambecco is an Italian instruction-tuned LLM (7B and 13B) built to study how model performance emerges from the interplay of data, compute, and optimization trade-offs.
How I approached it 🗺️
Developed end-to-end training pipelines, evaluation workflows, and inference tools using HuggingFace, PEFT/LoRA, and Gradio to systematically explore fine-tuning dynamics.
Why it matters 💡
Reproducible notebooks and a user-friendly Gradio interface that enable researchers to experiment with Italian-language model adaptation.
Highlights 🌟
Key takeaways with the most relevant technical signals in focus.
1
Italian instruction-tuned LLM built from LLaMA foundation 2
PEFT/LoRA for memory-efficient fine-tuning on GPU hardware 3
End-to-end research infrastructure: training, evaluation, and inference 4
Reproducible notebooks combined with interactive Gradio interface 5
Systematic exploration of data quality, model scaling, and inference efficiency Technology stack 🧰
Grouped by role so the stack is easier to read at a glance.
Languages
🐍 Python
Frontend framework
🖥️ Gradio
Infrastructure
📓 Colab
AI/ML
🦙 LLaMA 🤗 HuggingFace 🔧 PEFT/LoRA 🧠 TensorFlow
Project profile fit 🧠
A compact view of the signals this project communicates.
What it reflects ✨
Ability to learn complex topics fast
Thinking in trade-offs
Quantitative and experimental mindset
Understanding of ML economics
What it shows 🔎
AI/ML expertise
Model training
Research mindset
Product management
Open the live project to explore the full implementation and documentation.
Open Stambecco 🌍