Alconost drops pro-grade MQM gold dataset

This is the kind of small-but-clean dataset release that can punch above its size in MT eval workflows.

• –The data includes full MQM structure (category, severity, span) plus multiple annotators per segment, which is unusually useful for agreement and metric-analysis work.
• –Reported Kendall’s τ = 0.317 is materially above typical WMT ranges cited by the authors, suggesting annotation process quality was a priority.
• –The dataset is better suited for benchmarking, error analysis, and reward-model calibration than for training large translation models due to its scale.
• –License is CC BY-SA 4.0 and the card notes it is a growing collection, so it could become a recurring reference set if updates continue.
• –If replicated by others, this could raise expectations for professionally annotated open MT eval resources.