The student worked with literature recommended by the supervisor and independently searched for additional sources.

The written part of the thesis was prepared during the last month, but progressively, and I was gradually familiarized with its content. The student incorporated feedback, and the completion of the thesis proceeded without significant time-related complications. The final content was sufficiently consulted.

The technical report is of high quality. It starts with an introduction to artificial intelligence, language modeling and Large Language Models (LLMs). This is followed by an overview of modern parameter-efficient fine-tuning techniques and mathematical reasoning models. Then, it describes the chosen training dataset, GSM8K, three LLMs to be fine-tuned, and the selected fine-tuning strategy based on low-rank adaptation (LoRA). It follows with implementation details and experimental results, where the performance gains of fine-tuned variants of the chosen models are discussed. The final chapter contains the conclusion.

I was missing more theoretical background on the specific thesis topic. The thesis dives unnecessarily deep into AI/ML basics and then misses some key definitions, which are later referenced in the text, despite the lack of detailed explanation. This includes, for instance, references to multi-head attention in Section 2.4 and to Reinforcement Learning with Human Feedback and Proximal Policy Optimization in Section 4.3. In general, I expected the thesis to delve a bit deeper into mathematical reasoning and different fine-tuning strategies. Finally, figures such as Figure 2.2, could use more detailed descriptions.

The student selected a suitable dataset and fine-tuned three different LLMs using LoRA with a grid-like hyperparameter search. The experimental evaluation and discussion are sound, comparing different hyperparameters and model variants against each other.

However, significant effort focuses on hyperparameter tuning, even though the tuning methodology is rather simple. This translates to the thesis itself, which in general feels a bit minimalist. While the student clearly demonstrated his ability to perform and discuss the results of relevant experiments, he only applied the same methods on the same dataset to different pre-trained models. I was missing a comparison of different fine-tuning strategies (some of which, such as QLoRA, the student mentions himself), different math datasets, and a clearer emphasis on chain-of-thought reasoning.

Programmatically, the solution is a collection of Python scripts, which are well-structured but largely undocumented.

Evaluation level: zadání splněno

The assignment was fulfilled completely and all goals were sufficiently addressed.

The thesis cites a total of 33 sources, mostly scientific articles but also a significant amount of books. The sources are highly relevant, recent, and correctly used and cited in the text. I am not aware of any major violations of citation ethics. However, I would expect direct citations of all referenced methods, including older ones such as LSTM and GRU in Section 2.4.