You can find in this repo a brief notebook of the final model that we used in Defi-IA 2021 competition.
This edition of the Défi IA pertains to NLP. The task is straightforward: assign the correct job category to a job description. This is thus a multi-class classification task with 28 classes to choose from.
The data has been retrieved from CommonCrawl. The latter has been famously used to train OpenAI's GPT-3 model. The data is therefore representative of what can be found on the English speaking part of the Internet, and thus contains a certain amount of bias. One of the goals of this competition is to design a solution that is both accurate as well as fair, as explained in the Evaluation section.
You can find a full description here : https://www.kaggle.com/c/defi-ia-insa-toulouse
We had the oppportunnity to be at the second place for the fairness metric and the 4th place for f1 score .