Part 1 is done. This page recaps what you can now do, checks it with five questions, and lists what Part 2 assumes.

What you can do now

1. Explain next-token prediction, tokens, context window, parameters and quantization in plain words, and connect them to file size, memory and speed.
2. Place a model on the four-tier openness ladder and say why libraries care (provenance, reproducibility, vendor independence).
3. Install Ollama or LM Studio, adjust the privacy settings, download a model that fits your hardware, and check that it works.
4. Write prompts for cataloging-style tasks and improve them step by step.
5. Recognize a confident wrong answer and know two checks for it.

Self-check

Answer in your head or on paper first, then open the answer.

1. Why can a model give you 21 words when you asked for exactly 20?

Show the answer

It predicts one token at a time and has no counter, so it follows length instructions only approximately.

2. An 8B model file is about 5 GB at Q4_K_M. How much memory does it need, and what adds to it?

Show the answer

At least the file size, because the whole file is loaded into memory. The context window adds more: for the 8B model, each token of context costs 128 KiB.

3. Your prompt asks for books before 1950, sorted by year. The answer is tidy but wrong. What do you change?

Show the answer

Split the task. First ask the model to extract every book with its year. Then filter and sort in a second step, or do that part without the model. Check the result against the text.

4. Which tier of the openness ladder is Apertus, and why is that not the same as open source?

Show the answer

Tier 3, fully open: the weights, training code and checkpoints are published, with scripts to rebuild the training data. Open weights alone puts a model on a lower tier. The Open Source Initiative (OSI) definition of open source asks for information about the training data. Llama 2, for example, fails it although its weights can be downloaded.

5. The model gives you an ISBN, a DOI or a thesaurus term. Name two checks.

Show the answer

Look it up in a real source: a library catalog such as the DNB catalog for an ISBN, Crossref for a DOI, or the thesaurus itself for a term. Ask the same question twice in new chats and compare the answers.

What Part 2 assumes

Part 2 runs notebooks on your machine. Before the break, tick these:

• ollama pull nishad/apertus:8b (or nishad/apertus:4b on an 8 GB machine) has finished.
• ollama pull qwen3-embedding:0.6b has finished.
• ollama ps shows the model while it is loaded.
• Python 3.12 or 3.13 is installed.
• uv is installed.
• The notebooks repository is cloned.
• 00_check_setup is green on every line.

The full steps are on Am I ready?. If any box is not ticked, use the break to fix it, or ask for help. Part 2 starts with the environment check, described on Environment.

Leaving now?

You can stop here. Part 1 stands on its own. Everything in Part 2 is available afterwards as executed notebooks with their outputs, in the notebooks repository: github.com/nishad/llm-workshop-notebooks. Read them at your own pace, and run them when you are ready. Part 2 describes what they cover.