Each team had access to train, and validation XML sets of parallel sentences in French and Bambara as a baseline. The test set was hidden and used to evaluate the final scores. My group's approach got us to 2nd place in this challenge. We explored various approaches and combinations of parameters for training models, but as usual in these kinds of tasks, what gave us the edge over other groups was better data mining and processing that got us access to unique training data. This is critical for said "low-resource" languages like Bambara. The paper is public for more details.