This page contains information and teaching materials for the “LLM” lecture (for the year 2025-2026)

The aim of this course is to give you an overview of how LLM works and the different possibilities offered by these new models. We will see how to apply these models (as well as more traditional approaches) to a range of NLP tasks, from classification and machine translation to machine transcription and generation.

All lab assignments must be sent directly to my email address and must not exceed the deadline (any lab submitted after the deadline will not be marked). You must submit a report in PDF format; the file name must include your surname, and the header of the report must state your degree programme. Only include in the lab report the answers to the questions and the code that was used to obtain those answers.

Policy regarding the use of AI tools such as ChatGPT:

I have no objection to you using generative AI tools to assist you with your lab reports (LLMs are, after all, one of the main subjects of this course, and I have no difficulty admitting that these very lines will probably pass through a generative AI between the moment they leave my fingers and the moment they reach your retina).

That said, I would like to draw your attention to the fact that, like any tool, AI must be used properly, and poor use is generally more harmful than helpful (we will return to this issue at length during the course). In particular, several studies suggest that using AI during the learning process is not necessarily a good idea (to put it mildly), and that one must be able to answer questions without AI before being able to answer them with AI.

I will assume that you are adults, responsible for your choices, and aware that you will one day have to sit job interviews in which you will be expected to demonstrate that you can code, reason, and solve problems on your own and without AI (though I may of course be mistaken on this point).

I will simply ask you to indicate whether you used an AI tool to generate code or elements of your answers. This will not affect your mark, only the amount and type of feedback I provide.

Lecture n°1 (2026-01-23)

  • Reminder: Linear Classifiers and Learning Theory [pdf]
  • Lab n°1 [pdf] due date: February 13th
  • data set for the langid [full data] (21Go uncompressed) [large] (1.7Go uncompressed) [small] (173 Mo uncompressed)

Lecture n°2 (2026-01-30)

  • no lecture this week 😭

Lecture n°3 (2026-02-06)

  • Quiz n°1: The test will cover everything that was discussed during the previous lecture (at least as far as I remember) as well as all the material included in the slides, up to and including slide 48 (i.e. the section on feature extraction). It consists of short, simple, and precise questions (according to my definition), typically between 10 and 30 questions depending on my inspiration at the time.

No documents, AI tools, friends, pets, or any other form of assistance are allowed, including but not limited to external expertise, collective intelligence, benevolent spirits, or unexpected flashes of shared enlightenment.

Lecture n°4 (2026-02-12)

  • Quiz n°2
  • lecture: Word Representations: word2vec & Transfomers [pdf]
  • Lab pdf tsv － due date: March 6th the lab has been update (Feb. 23rd)

Lecture n°5 (2026-02-19)

  • Quiz n°3
  • lecture: Transformer, BERT & HuggingFace API (same slide as previous lecture)

Lecture n°6

  • new lab 🥳🥳 pdf
  • Transformer: the details [pdf] (slides are not polished yet 😢)

LEcture n°7

  • Generating text with Transformer: the decoder architecture [pdf]

Lecture n°8 (2026-03-27)

  • From LMs to LLMs [pdf]

Lecture n°9

  • Audio models [pdf]

Lecture n°10

Cultural References (For the Historically Curious)

Because so many of you have asked (well, strictly speaking, more than two), and because it is, after all, my solemn duty as a lecturer to broaden your cultural horizons and gently remind you that remarkable things did in fact exist before the year 2000, here are links intended to “explain” some of the cultural references (selected, of course, with impeccable taste, undeniable flair, and unwavering relevance) that occasionally appear in my slides:

Results

  • all corrected labs are here