metadata
dataset_info:
features:
- name: sentence1
dtype: string
- name: sentence2
dtype: string
- name: labels
dtype: int64
- name: metadata
dtype: string
splits:
- name: train
num_bytes: 12201493
num_examples: 27422
- name: test
num_bytes: 3190343
num_examples: 6589
download_size: 5901103
dataset_size: 15391836
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
license: other
source_datasets:
- irds/antique
- irds/antique_train
- irds/antique_test
antique
All native ANTIQUE judged pairs, retaining relevance grades 1–4.
Join the original query, answer-collection and qrel files without generated negatives or filtering grades. Preserve native query-disjoint train/test. Q0/E0/U0 judgment provenance stays in metadata. The original readme permits research use; Yahoo answer rights are not replaced by a software license. Original distribution: https://ciir.cs.umass.edu/downloads/Antique/
Original data: irds/antique, irds/antique_train, irds/antique_test, original distribution. Repackaged as parquet for tasksource by scripts/repackage_dataset/relevance.py.
Conversion provenance: source revisions, hashes, configuration and exclusions.