Datasets:
metadata
annotations_creators:
- found
language_creators:
- found
language:
- en
license:
- unknown
multilinguality:
- monolingual
size_categories:
- 10K<n<100K
source_datasets:
- original
task_categories:
- summarization
task_ids:
- summarization-other-paper-abstract-generation
paperswithcode_id: multi-xscience
pretty_name: Multi-XScience
This is a copy of the Multi-XScience dataset, except the input source documents of its test
split have been replaced by a dense retriever. The retrieval pipeline used:
- query: The
related_work
field of each example - corpus: The union of all documents in the
train
,validation
andtest
splits - retriever:
facebook/contriever-msmarco
via PyTerrier with default settings - top-k strategy:
"max"
, i.e. the number of documents retrieved,k
, is set as the maximum number of documents seen across examples in this dataset
Retrieval results on the test
set:
ndcg | recall@100 | recall@1000 | Rprec |
---|---|---|---|
0.3783 | 0.5229 | 0.7217 | 0.2081 |