Ph.D. thesis of Kamyar Zeinalipour, defended on 10 February 2026 at the University of Siena (Ph.D. in Information Engineering and Science), supervised by Marco Gori. Grade: eccellente.
Abstract
The thesis builds a framework for turning general-purpose large language models into small, open educational tools, with a focus on languages that have little training data. A large model first turns raw text into structured teaching material; the result, a family of instruction datasets called the Instruct series, is then used to fine-tune smaller open models for one learning task each.
The framework is applied to a sequence of tasks: educational crossword puzzles in English, Italian, Arabic and Turkish; multiple-choice and short-answer questions in Persian and Turkish; and automated syntax feedback on student essays in English. Human and automatic evaluations in each language show that the smaller fine-tuned models become reliable specialised tools.
The work produced 11 publications, 11 datasets (7 of them in the Instruct series) and 29 fine-tuned generator models across 5 languages. The code, data and models are public: see Language models and datasets.