The forward translation model was trained on a curated, synthetic dataset and tested on a 1k test set translated by native speakers. The back translation model switched the source and target languages, proceeding with the same data but in reverse.
Forward: English -> Dagaare res[test] = {'chrf': 22.10890708580873, 'bleu': 1.2166438958332253}
Backward: Dagaare -> English {'test': {'chrf': 22.477936794013537, 'bleu': 4.076082383321696}}
