Lecture slides, notebooks, and datasets from every edition since 2023. Openly available, no registration required.
Watch this space. Materials go up as speakers clear them for release, so a session missing today may well be here next month. We are also editing the lecture and research talk recordings and will publish them openly once post-production is done.
Agents and how to evaluate them, continual learning and why models forget, and how training data, evaluations, and RL environments shape model behaviour. Efficiency runs through the year: compression, self-hosted serving, and the economics of inference. The challenge materials add a survey-prediction dataset and a working baseline pipeline.
Foundations you can run: tokenization, chat templates, decoding strategies, and attention, then post-training end to end, from supervised fine-tuning through preference optimisation and GRPO, with datasets and training code included.
Slides are being collected from speakers and will appear here.
Model architecture, evaluation, and agent-based systems, with fine-tuning and alignment covered in depth. Applied sessions turn to retrieval-augmented generation for social science, observability for LLM pipelines, and the practicalities of self-hosting.
Transformers from the ground up, interpretability and bias, and alignment. The notebooks remain a solid starting point: topic modelling, text classification, parameter-efficient fine-tuning, model analysis with spaCy, and detoxification with RLHF.
All lectures by Elena Voita.
All workshops by Ilya Boytsov.