webfaq_cleaned

A cleaned, multilingual question–document dataset across 44 languages, ~58.5M Q–D pairs total, intended as a contrastive training / retrieval-eval corpus. Derived from PaDaS-Lab/webfaq.

The source data is web-scraped FAQ pages. The raw input contained large amounts of SEO doorway pages, casino spam, e-commerce listing pages whose review snippet didn't match the product query, near-duplicate templates, and broken encodings. This dataset applies a multi-layer cleaning pipeline to filter that noise while preserving the substantive Q&A.

Schema

Each row is a single Q–D pair:

field	type	description
query	string	the question
document	string	the answer / supporting passage

One config per language; one train split per config.

Usage

from datasets import load_dataset

# Load one language
ds = load_dataset("bowang0911/webfaq_cleaned", "en", split="train")
print(ds[0])
# {'query': '...', 'document': '...'}

# Or stream it
ds = load_dataset("bowang0911/webfaq_cleaned", "ja", split="train", streaming=True)
for row in ds.take(5):
    print(row)

Languages

44 languages, sorted by final row count. source = rows pulled from the original PaDaS-Lab/webfaq shards; kept = rows remaining after cleaning.

lang	source	kept	kept%	lang	source	kept	kept%
en	49,275,346	40,817,655	82.8%	ro	546,709	378,587	69.2%
fa	969,748	841,712	86.8%	fi	641,009	373,519	58.3%
ar	1,000,000	829,000	82.9%	th	502,215	355,047	70.7%
tr	1,000,000	820,040	82.0%	el	494,019	349,025	70.7%
fr	1,000,000	805,753	80.6%	hu	363,866	283,635	78.0%
nl	1,000,000	802,485	80.2%	no	316,320	248,062	78.4%
de	1,000,000	798,319	79.8%	he	391,084	188,530	48.2%
es	1,000,000	797,885	79.8%	sk	177,748	141,875	79.8%
it	1,000,000	796,614	79.7%	bg	164,015	134,795	82.2%
pt	1,000,000	793,637	79.4%	lt	119,192	89,490	75.1%
ru	1,000,000	784,173	78.4%	is	91,630	64,358	70.2%
pl	1,000,000	782,624	78.3%	lv	79,443	59,446	74.8%
vi	1,000,000	723,073	72.3%	ca	80,189	57,740	72.0%
zh	1,000,000	682,722	68.3%	et	69,643	56,385	81.0%
sv	832,140	675,465	81.2%	bn	57,943	50,329	86.9%
ja	1,000,000	662,302	66.2%	sl	49,867	42,019	84.3%
da	768,688	609,387	79.3%	mr	30,658	26,642	86.9%
ko	646,996	592,344	91.6%	ms	24,448	20,867	85.4%
uk	805,505	580,032	72.0%	te	23,949	19,578	81.7%
id	652,660	482,649	74.0%	az	19,738	18,357	93.0%
hi	537,488	478,818	89.1%	tl	21,467	16,743	78.0%
cs	538,343	392,826	73.0%	ta	19,815	16,305	82.3%

Cleaning

Filters were developed iteratively: sample random rows per language, have an LLM inspect them to identify recurring failure patterns (templated SEO doorways, casino doorway pages, mismatched product-listing Q–Ds, encoding glitches), then encode those patterns as cheap substring/regex filters and re-sample to verify. Each row passes through four lightweight passes — a substring blocklist (cross-language SEO brand markers like KAYAK / momondo / Tripadvisor, plus a few per-language pattern sets added where sampling revealed language-specific noise ecosystems), a yes/no echo + query-template filter for the affected languages, near-duplicate dedup via MinHash LSH (64 perms, 16 bands, char-5gram shingles, Jaccard ≥ 0.85 per bucket), and encoding repair (HTML entities, JSON-unicode-escape leftovers, and literal \r\n\t fixed in place; rows still containing \ufffd after fix are dropped). The pipeline is deliberately conservative — it targets templated noise and known encoding glitches rather than scoring semantic relevance, so well-aligned Q–Ds are preserved even when the surface form is somewhat templated.

Known limitations

• Languages where product-listing SEO is heavy still have a residual ~15–20% mismatched pair rate after cleaning. The dominant failure mode is product-title queries paired with unrelated user reviews — each pair is unique (so dedup doesn't catch it) and lacks distinctive substring markers, so removing it cleanly requires a semantic Q–D relevance classifier rather than substring filtering.
• Languages with large kept% drops (he, fi, ro, th, vi, el, cs, etc.) reflect heavy removal of templated travel-aggregator hotel-CTA pages, not loss of substantive QA. Spot checks on the kept rows for these languages show 90–100% substantive content.
• Other languages (en, ar, hi, ko, de, fr, es, it, pt, ru, etc.) spot-check at 90–100% substantive. The remaining ~5–10% is mostly benign templated single-source answers (price predictions, weather summaries, hotel facts) rather than truly unaligned pairs.

Source & citation

This dataset is a derived, filtered version of PaDaS-Lab/webfaq. If you use this dataset, please cite the original WebFAQ release as well as this cleaning pass.

Intended for research use.