Large Language Models and Generative AI for NLP (University of Helsinki, Autumn 2026)
by University of Helsinki (Aarne Talman, Jussi Karlgren, Dmitry Kan)
A seven-week master's course with public slides and labs that takes you from a tiny GPT to a tool-using agent.
Overview
Large Language Models and Generative AI for NLP is a 5 ECTS course for master's students in language technology and data science at the University of Helsinki, listed as LDA-T318 in the 2026-27 LingDig teaching programme. It runs from 27 October to 10 December 2026. Aarne Talman teaches weeks 1 to 3 and 7, Jussi Karlgren teaches evaluation in week 4, and Dmitry Kan teaches RAG and agents in weeks 5 and 6. Each week has a Tuesday lecture and a Thursday lab: language models and pretraining (Lab 1: from text to a tiny GPT), post-training (Lab 2: does fine-tuning beat prompting?), inference (Lab 3: serve and measure your model), evaluation (Lab 4), retrieval-augmented generation (Lab 5: RAG for the project task), agents (Lab 6: build a tool-using agent), and a recap with project support. The lectures cover tokenization, perplexity, scaling laws and distributed training; chat templates, SFT, preference tuning, verifiable rewards and LoRA; decoding, KV-cache memory, batching, quantization and serving; benchmarks, contamination, model-based and multilingual evaluation; reranking and context assembly; and agent loops, memory and failure modes. Labs ship reference implementations and a local smoke test, and the GPU work is set up for CSC's Roihu cluster with Slurm. Assessment is a single project report. When checked on November 10, 2026, materials for weeks 1 to 3 were public and weeks 4 to 7 were still placeholders.
At a Glance
- Topic
- ML
- Level
- Intermediate
- Format
- Course
- Cost
- Free
- Duration
- 7 weeks, 27 October to 10 December 2026; one 90-minute lecture and one 90-minute lab per week, plus a final project
- Provider
- University of Helsinki (Aarne Talman, Jussi Karlgren, Dmitry Kan)
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Explain how decoder-only language models work, from tokenization and attention to next-token prediction
- ✓Train a tokenizer and a small GPT-style model from raw text in the first lab
- ✓Compare prompting against post-training methods such as SFT, preference tuning and LoRA on one task
- ✓Serve a model and benchmark it while reasoning about decoding, KV-cache memory, batching and quantization
- ✓Evaluate models with benchmarks and model-based judges while checking for contamination and multilingual variation
- ✓Build a RAG pipeline with retrieval, reranking and context assembly, then evaluate the RAG system
- ✓Build a tool-using agent and reason about agent loops, context management, memory and failure modes
- ✓Run ML workloads on an HPC cluster using modules, Slurm batch jobs and interactive GPU sessions
Highlights
- •Every lecture has a matching lab, and each lab ships a reference implementation plus a local smoke test you run before using a GPU
- •Covers the full stack in seven weeks: pretraining, post-training, inference, evaluation, RAG and agents
- •Evaluation gets its own week with a separate instructor, including contamination and multilingual evaluation that short courses usually skip
- •Slides and labs are published week by week in a public GitHub repo, so you can follow the live autumn 2026 run
- •The lab 'does fine-tuning beat prompting?' gives you a measured answer on your own task rather than a rule of thumb
Who It’s For
Best For
- ✓Engineers with Python and basic ML who want a structured university-level path across the whole LLM stack
- ✓NLP and data science students looking for public lab material on post-training, serving and evaluation
- ✓Practitioners who want to compare fine-tuning with prompting and measure RAG and agent quality themselves
- ✓Self-learners who want to follow a live course week by week, with the caveat that later weeks are still being published
Prerequisites
- •Python programming
- •Basic machine learning concepts
- •Familiarity with a Unix command line and git
- •PyTorch experience is useful but not required; outside learners need their own GPU, since the course GPU setup uses an enrolled-students CSC project
FAQ
What is Large Language Models and Generative AI for NLP (University of Helsinki, Autumn 2026)?
This is the University of Helsinki's autumn 2026 master's course on large language models, with slides and runnable labs published on GitHub as the course runs. It suits engineers who know Python and basic machine learning and want one guided path through pretraining, post-training, inference, evaluation, RAG and agents, each paired with a lab.
Is Large Language Models and Generative AI for NLP (University of Helsinki, Autumn 2026) free?
Large Language Models and Generative AI for NLP (University of Helsinki, Autumn 2026) is free to access.
What level is Large Language Models and Generative AI for NLP (University of Helsinki, Autumn 2026) for?
Large Language Models and Generative AI for NLP (University of Helsinki, Autumn 2026) is aimed at a intermediate audience. Recommended background: Python programming, Basic machine learning concepts, Familiarity with a Unix command line and git, PyTorch experience is useful but not required; outside learners need their own GPU, since the course GPU setup uses an enrolled-students CSC project.
How long does Large Language Models and Generative AI for NLP (University of Helsinki, Autumn 2026) take?
Expect roughly 7 weeks, 27 October to 10 December 2026; one 90-minute lecture and one 90-minute lab per week, plus a final project. Most learners work through it at their own pace.
What will I learn from Large Language Models and Generative AI for NLP (University of Helsinki, Autumn 2026)?
You'll learn: Explain how decoder-only language models work, from tokenization and attention to next-token prediction; Train a tokenizer and a small GPT-style model from raw text in the first lab; Compare prompting against post-training methods such as SFT, preference tuning and LoRA on one task; Serve a model and benchmark it while reasoning about decoding, KV-cache memory, batching and quantization; Evaluate models with benchmarks and model-based judges while checking for contamination and multilingual variation; Build a RAG pipeline with retrieval, reranking and context assembly, then evaluate the RAG system; Build a tool-using agent and reason about agent loops, context management, memory and failure modes; Run ML workloads on an HPC cluster using modules, Slurm batch jobs and interactive GPU sessions.
Topics
Sources
This page was written from 2 sources, 1 on domains other than github.com.