PaDaS-Lab/webfaq-ir-datasets

WebFAQ for ir_datasets

Archives used by the webfaq/* datasets in ir_datasets. They are downloaded automatically by ir_datasets.load("webfaq/<lang>/train") and similar; you normally do not need to fetch them by hand.

Each <lang>.tar.gz contains:

webfaq/<lang>/documents.jsonl        # {"doc_id", "text"}: answers, deduplicated by exact text
webfaq/<lang>/training/queries.jsonl # {"query_id", "text"}
webfaq/<lang>/training/qrels         # TREC qrels, one confirmed answer per question
webfaq/<lang>/training/BGE-M3.res    # TREC run: BGE-M3 teacher scores for positives + hard negatives
webfaq/<lang>/test/queries.jsonl
webfaq/<lang>/test/qrels

They are built from PaDaS-Lab/webfaq-retrieval (queries, answers, train/test splits) and IrvinTopi/WebFAQHardNegatives (BM25-mined hard negatives scored with BGE-M3). Answers with identical text are merged, keeping the first occurrence, and qrels and scored documents are remapped to the retained document IDs. The BGE-M3 scores are training signals for distillation, not relevance judgments.

Language	Documents	Train queries	Test queries	Train scoreddocs	Archive	MD5
ara	141,089	132,664	10,000	6,336,393	98 MB	fb1c59ccaba2221b1670b846aec77d59
dan	136,939	127,686	10,000	6,123,062	91 MB	218d6b3e8b4fbdfb759f4f79ec580319
deu	886,794	881,201	10,000	18,702,436	550 MB	a0d76a212bdbff0839013f8375c7a562
eng	5,252,530	5,268,725	10,000	25,632,193	1,957 MB	14611295d1c8eedd80db4ca15f5f90e1
fas	226,446	216,940	10,000	12,654,983	219 MB	1921d193ef4cfd90ae5fba5974b4ca7a
fra	565,142	559,505	10,000	19,140,183	434 MB	77f5bce88f2c84485e168146bda3a553
hin	99,897	90,031	10,000	6,374,233	101 MB	738f7ba66240833dc799f6ef946d6956
ind	109,774	101,315	10,000	6,134,393	99 MB	6618692cb8bc84f22c091cba4eb8d2a1
ita	255,534	247,803	10,000	18,887,103	310 MB	4f89ed3a74af5d19353f8764de5260ff
jpn	307,368	299,157	10,000	19,111,766	344 MB	fe4f4d8d1b2df9a681273c59a37ec2ef
kor	101,036	92,000	10,000	6,347,985	103 MB	10368db40e0f83eb6a51409aa8dc4d14
nld	368,382	360,662	10,000	18,627,158	343 MB	645cee39c4ed725fb178f1a1d7060a82
pol	191,578	182,515	10,000	12,775,737	213 MB	a922476ed3754e26ee818031a2f1f131
por	208,209	199,353	10,000	12,731,380	215 MB	6dddcf33d28117cdfcdd33117b7f95af
rus	383,269	377,504	10,000	19,040,146	389 MB	0527dba5764be2aada5be8fa7cd1febe
spa	600,855	594,661	10,000	19,015,957	433 MB	2cbbc80199bdfd82578c7b94a97ae0dd
swe	157,849	148,738	10,000	6,155,553	116 MB	c5a6c5195a746908b34fde4c73de8935
tur	144,115	134,846	10,000	6,327,043	112 MB	53f013906712957365d875ff60c63bbc
vie	123,283	113,972	10,000	6,401,341	116 MB	70a4f08bd733d89eb799cf8fcbbe3f39
zho	131,606	122,491	10,000	6,379,527	120 MB	55477ad01d5dd79fee38ef56ed4e0e2c

Citation

@inproceedings{Dinzinger2025WebFAQ,
  title = {{WebFAQ}: A Multilingual Collection of Natural {Q\&A} Datasets for Dense Retrieval},
  author = {Michael Dinzinger and Laura Caspari and Kanishka Ghosh Dastidar and Jelena Mitrovi{\'c} and Michael Granitzer},
  booktitle = {Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  series = {SIGIR '25},
  year = {2025},
  doi = {10.1145/3726302.3731934}
}

@inproceedings{Dinzinger2026WebFAQ2,
  title = {{WebFAQ} 2.0: A Multilingual {QA} Dataset with Mined Hard Negatives for Dense Retrieval},
  author = {Michael Dinzinger and Laura Caspari and Ali Salman and Irvin Topi and Jelena Mitrovi{\'c} and Michael Granitzer},
  booktitle = {Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  series = {SIGIR '26},
  year = {2026},
  doi = {10.1145/3805712.3808621}
}