As the use of generative AI such as ChatGPT increases, so does the associated global resource consumption, to the point where Big Tech is reportedly considering operating fault-prone nuclear reactors. However, for many tasks, alternative, less resource-intensive approaches often achieve similar or even better performance than large general-purpose language models. Born out of the need to classify CorrelAid projects according to the Sustainable Development Goals (SDGs) they advance, this coding session aims to strike a balance between performance and resource consumption of multi label text classification methods.
We will start with a reflection on the current developments in AI usage and their consequences, including the ethical implications for Data4Good. A tutorial on multi-label text classification using binary relevance with TF-IDF and Logistic Regression follows. We will also introduce Modal for serverless code execution and the codecarbon package for measuring carbon emissions. The session will conclude with a coding session where participants can practice the introduced methods through predefined tasks or engage in freestyle exploration.
-
Install Poetry
Follow these instructions.
-
Fork or clone this Repo
-
poetry install -
Create an account on modal.com
-
Run
poetry run modal setup
poetry run modal run scripts/train_model.py
-
pandoc -t revealjs -s -o documents/presentation/slides.html documents/presentation/slides.md -V revealjs-url=https://unpkg.com/reveal.js/ --include-in-header=documents/presentation/slides.css