COMET

Why we included this project

COMET exists to answer a question that comes up constantly in machine translation work: is this new MT system actually better than the one we already run? It scores translations against human references, and in referenceless mode it works from just the source and output text, which covers the common case where no gold translation exists. The framework also lets you train your own evaluation metrics on your own data, so it works as both a ready-made scorer and a research toolkit. The newer XCOMET models add error-level detail, flagging which errors in a translation are minor, major, or critical instead of returning a single quality number. If you evaluate MT systems, benchmark models, or build translation tooling, the CLI and docs make it easy to get started.