This repo contains the synthetic data for the Nugget paper.

The file docs.txt contains all the documents. Each line contains a doc id and the doc itself, separated by \t.

The file task.jsonl contains the metadata of the task. Each line is a data point, where source is the id of the source doc, candidates are candidate doc ids, and the answer is the index of the target document in the candidate list.

Note that the paraphrase identification dataset contains 3 splits, but we only experiment with the dev set.