Translating games with small models
Whether models small enough to run on an ordinary machine can translate game dialogue, including into low-resource languages.
Text-heavy games live or die on their writing, and official localization rarely reaches beyond the commercially obvious languages. We are evaluating accessible models on game dialogue across several target languages, combining native-speaker judgement with automatic measures, and paying particular attention to the things that break in practice: terminology drift, tone, and the variables and tags embedded in the text.