AlignScore

AlignScore is a metric for automatic factual consistency evaluation of text pairs, determining whether a context supports a claim and whether the claim contradicts the context. It is applicable to tasks like summarization, paraphrase, and dialogue, and is evaluated on a leaderboards comparing AlignScore to other metrics. The project provides two RoBERTa-based checkpoints, AlignScore-base and AlignScore-large, with installation and usage guidance to compute scores via PyTorch on the Hugging Face hub.

AI named AlignScore from August 2026 to September 2026.