Ostr-AI 2026 – AI Summer School 2026

Overall Goal

Ostr-AI logo The Ostr-AI 2026 is designed to provide participants with both a solid theoretical understanding of modern Large Language Models (LLMs) and hands-on experience in developing practical AI solutions. The central objective of the programme is to guide participants through the implementation of a non-trivial AI project in Google Colab, focusing on the fine-tuning of an LLM for a domain-specific task.

Each participant will gain practical experience with advanced LLM fine-tuning techniques. The project will be introduced during the two-day on-site session, where participants will receive all necessary datasets, resources, and methodological guidance. By combining the skills acquired during the lectures and practical sessions, participants will learn how to build efficient and cost-effective AI systems capable of processing and analyzing large volumes of textual data.

A significant part of the programme is devoted to independent project work. After several weeks of work participants will present their results during an online follow-up session. Participants will receive support and feedback from members of the research team at IRAFM, who are actively involved in research on large language models and artificial intelligence.

Venue

Institute for Research and Applications of Fuzzy Modeling
University of Ostrava, Bráfova 7, Ostrava
Contact e-mail:

Dates

  • On-site session: 26–27 August 2026
  • Online project presentations: September 2026 (exact date to be announced)

Participants

The summer school is intended for a selected group of approximately 15–20 participants, ensuring an interactive learning environment and sufficient individual support.

Target Audience and Requirements

The programme is suitable for:

  • Master's and doctoral students in Mathematics, Computer Science, Data Science, and related disciplines;
  • Early-career researchers and professionals interested in applying modern AI methods in their work.

Researchers from a broad range of disciplines are encouraged to apply, including:

  • Medicine and Healthcare
  • Economics and Business Sciences
  • Social Sciences
  • Psychology and Behavioural Sciences
  • Natural Sciences and Engineering
  • Law and Public Administration

Basic programming skills are recommended. Prior experience with Python or machine learning is beneficial but not mandatory.

Registration and Selection

Ostr-AI logo Registration is available via the online application form.

As the number of places is limited, participants will be selected by the organizing research team. Applicants will be asked to provide:

  • their area of expertise and research interests,
  • their previous experience with AI and large language models,
  • their level of Python programming proficiency.

Application Timeline

  • Registration deadline: 10 July 2026
  • Notification of acceptance: within several days after the registration deadline

Participation Fee

Participation in the Ostr-AI 2026 is free of charge.

Detailed Programme:

1.
Large Language Models (60 min) (P. Sosík)
  • Transformer architecture: encoder/decoder blocks, tokenization, positional encoding, sequence processing
  • Self-attention: queries, keys, values; multi-head attention; pattern visualization
  • Scaling laws and emergent capabilities
2.
Data Preparation (60 min) (P. Hyner)
  • Text cleaning, normalization, deduplication, handling noisy real-world data
  • Batching, padding, attention masking, efficient DataLoader pipelines
  • Data quality impact on model performance
  • Hands-on messy dataset exercise
3.
Intro to PyTorch (60 min) (D. Mojžíšek)
  • Tensor operations, autograd mechanics, basic neural network components
  • Model building blocks: layers, optimizers, loss functions, and training loops
  • Writing and debugging a full training loop from scratch
4.
HuggingFace (60 min) (D. Adamczyk)
  • Loading datasets and pretrained models, reading model cards and configs
  • pipeline(), Trainer API, tokenizer usage
  • Evaluations, benchmarks, Open LLM Leaderboard navigation
5.
Model finetuning (60 min) (K. Krmaschek)
  • Transfer learning: layer freezing, learning rate scheduling, model adaptation strategies
  • Parameter-efficient finetuning: LoRA, PEFT; instruction following; few-shot tuning
  • Knowledge distillation: teacher–student compression (optional)
6.
API models (60 min) (A. Zlámal)
  • OpenRouter overview; cost comparison: API vs. local inference vs. subscriptions
  • LLM-assisted data labeling and synthetic training data generation
  • Practical budgeting strategies for prototyping
7.
AI Agents, Tool Use & MCP (60 min) (J. Hůla)
  • Agentic LLM architectures: tool calls, memory, multi-step reasoning loops
  • Defining skills, chaining actions, handling failures
  • Model Context Protocol (MCP): connecting models to external services
  • Live agent demo
8.
Final project (60 min) (the whole team)
  • Intro, necessary resources, hints, techniques, and leaderboard.


Updated: 25. 06. 2026