I'd like to start by thanking @JEANMPIA on his contributions in the discussions. I can't recall how many times I said I was done with this contest :) I started working on it, 2 weeks before the end, and got stuck at 0.45/0.47 for the most part of it. Reading the discussions kept me going and ultimately led to this result.

Overview

My solution is an Afro-xlmr-large model finetuned on the provided training data, some additionnal data ( afrikaans, arabic, french and english ) and the 2nd rounds of pseudo labels done on Luo & Setswana monolingual data.

Pipeline

You can find the code here on github.