Tools & Methods

The lab maintains practical guidance on the methods its researchers use, including:

• Choosing where to run a model. Our routing rule, in order: can the data legally leave Bocconi; must the result replicate in five years; does a fine-tuned small model do the job; only then, cost.
• Encoder models for classification. When a fine-tuned BERT-family model beats a large language model on accuracy, cost and reproducibility.
• Open-weight models on Bocconi's computing cluster. Running, quantising and fine-tuning models on the university's GPUs.
• Reproducibility with LLMs. Recording models, versions, prompts and settings so results can be replicated.

Code templates and guides are shared with incubated teams and AI Clinic users.