Sentence-level attribution in multi-document summarization: human-annotated attribution set and replication package

This record accompanies the article "Sentence-level attribution in multi-document summarization: scorer sensitivity and calibration against human labels" (submitted to Language Resources and Evaluation). It contains a human-annotated attribution set of 150 summary sentences from the Multi-News test split. Two annotators labeled whether each sentence is supported by any of its retrieved source passages and selected the sentences that carry the support; 40 items were labeled by both. Each item also records the decisions of eight support-scorer configurations, so new scorers can be evaluated against the labels without new annotation. The record also holds the per-link model-judge verdicts at document and span granularity, all per-example scores behind the article's tables, the cached generations, every prompt, the annotation tool and guidelines, and the pipeline code. Every number in the article can be recomputed without a GPU: run replication/reproduce.sh inside replication-package.zip. README.md describes the files, and NOTICE.md lists the redactions and the terms that apply to the third-party news text. Data: CC BY 4.0. Code: MIT.