Develop episode 4: LLM

[laurasootes]

4 hours of lesson material

Lesson questions:

Lesson objectives:

Challenge to solve this episode: compare a newspaper from the fifties with one form the nineties. How do we speak about XXX in those different times?
Data: Delpher newspaper articles.

General activities:

• prompt engineering
• how do you implement a LLM in your research, what next steps can you do with a generated summary for example

Current idea: select two Delpher newspapers for the same specific date, but from different periods: one from the ‘30ies or ‘50ies (whichever works best) and one from the nineties. And compare the results for different tasks:

Steps:

• Collect the titles of the articles.
• Get a one-sentence description of the articles
• Classify the articles: e.g. politics, economics, sports, culture...
• Compare the writing style

Optional dates (decide which works best later). Newspaper itself to be decided (e.g. AD, Volkskrant, telegraaf, RNC...)

• 1 January
• Kings/queens day
• May 5th (oldest newspaper from the fifties)

Subjects to discuss:

• What is an LLM and how is it different from other NLP models
• How does an LLM work, maybe something on the architecture.
• Which LLMs are out there, how do you choose which one to use?
• Which tasks are LLMs good at? For which tasks are other models better?
• (how can you fine-tune a LLM model)
• How do you use an LLM such that you get the best results
• Important: pitfalls, limitations, caveats, privacy

resources:
https://github.com/meta-llama/llama-recipes/tree/main/recipes
https://research-software-directory.org/projects/navigating-stories
https://www.delpher.nl/nl/kranten/results?query=&coll=dddtitel&sortfield=date
https://github.com/HandsOnLLM/Hands-On-Large-Language-Models