Exercises (DL4NLP 1)

Exercise 1

Building, Tuning and Evaluating a standard RNN Model: Notebook

Exercise 2

Building the Vaswani et al. (2017) “Vanilla” Transformer from scratch: Notebook

Exercise 3

In this exercise, we will be using the Hugging Face ecosystem to finetune BERT for a classification task (part 1) and T5 for text summarization (part 2). Then, we will upload our training and validation results via TensorBoard, which makes interactive sharing and inspecting of results very easy: Notebook

Part 1 and part 2 can be worked on independently.

Exercise 4

In this exercise, we will first look at the inner workings of the attention mechanism and the importance of various heads in pretrained models. The second task compares multiple decoding strategies on the same model and prompt to highlight how much on an effect the decoding algorithm has on the quality of generated text: Notebook

Task 1 and Task 2 can be worked in independently.

Exercise 5

In this exercise, we will use Qwen3 to learn about and experiment with various prompting strategies to solve classification tasks with text generation models Notebook

Exercises (DL4NLP 2)

to be created/added

• »