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 -> Kanuri RunModel: res[test] = {'chrf': 21.84594377560411, 'bleu': 0.21784873577250005}
Backward: Kanuri -> English {'test': {'chrf': 30.433710037476953, 'bleu': 4.985695668743153}}
