mqm

This paper demonstrates that small open-weight LLMs (<30B parameters) can achieve competitive interpretable translation quality estimation, including MQM error annotations and corrections, rivaling much larger proprietary models while preserving data privacy.