PaDaS-Lab/webfaq-v2

WebFAQ 2.0: Multilingual FAQ Q&A Dataset with Hard Negatives

Overview | What's New in v2.0 | Dataset Statistics | Structure | Bitext Alignments | Hard Negatives | Training Strategies | Examples | Considerations | License | Citation | Contact

---

Overview

WebFAQ 2.0 is a large-scale multilingual dataset of 198 million natural question–answer pairs across 108 languages, mined from structured FAQ pages on the web.

It is the successor of the original WebFAQ (v1) dataset (96M QAs, 75 languages): 👉 https://huggingface.co/datasets/PaDaS-Lab/webfaq

WebFAQ 2.0 significantly expands multilingual coverage, bilingual QA alignments, and introduces a new hard negatives dataset for training dense retrieval models .

The dataset is designed for:

• 🌍 Multilingual & cross-lingual dense retrieval
• ❓ Open-domain Question Answering
• 🔎 Hard negative mining research
• 🌐 Bitext mining & multilingual embedding training

---

What's New in v2.0

Compared to WebFAQ v1, version 2.0 introduces:

1️⃣ Massive Scale Increase

• 198M QA pairs (vs. 96M in v1)
• 108 languages (vs. 75)
• English share reduced from ~51% to ~28% → stronger multilingual balance

2️⃣ New Crawling Strategy

Instead of relying solely on Web Data Commons dumps, WebFAQ 2.0:

• Mines FAQPage markup directly from Common Crawl

• Crawls pages using OWLer (OpenWebSearch project)

• Extracts:

  • Questions
  • Answers
  • Page titles
  • Page descriptions
  • hreflang links for multilingual alignment

This yields richer contextual metadata and improved cross-lingual coverage .

3️⃣ Large-Scale Bilingual Alignments (👉 michaeldinzinger/webfaq-v2-bitexts)

• 14.3 million aligned QA pairs (vs. 1.5M in v1)
• 3,970 language combinations (≥100 aligned samples each)
• 3,118 language pairs contain ≥1,000 and 1,282 language pairs contain ≥4,000 aligned QA pairs

This represents nearly a 10× increase in alignment size.

4️⃣ Hard Negatives Dataset (👉 IrvinTopi/WebFAQHardNegatives)

A new dataset of:

• 1.25M queries
• 20 languages
• 200 mined hard negatives per query
• Cross-encoder scores included

Designed specifically for dense retriever training .

---

Dataset Statistics

Languages

• 108 languages with ≥1,000 samples
• More balanced distribution compared to v1
• Significant growth in non-English languages (e.g., Hindi, Ukrainian, Dutch, Portuguese, Polish)

Topics

Topic labels are inferred using a fine-tuned XLM-RoBERTa classifier trained on GPT-5-mini annotated data (F1 ≈ 88%).

Major topic shift in v2.0:

Topic	%
✈️ Traveling & Hospitality	59.2
🛒 Products & Commercial Services	18.9
❤️ Healthcare & Lifestyle	5.6
🎵 Entertainment	4.7
🏦 Banking & Finance	4.4

Question Types

Questions are categorized using a multilingual extension of the Bolotova et al. taxonomy:

Type	%
Not-a-Question	34.7
Factoid	33.6
Experience	11.6
Instruction	10.3
Evidence-Based	4.3
Reason	3.3
Comparison	2.2
Debate	0.1

Classifier F1 ≈ 88%.

---

Structure

Each language is provided as a separate configuration:

from datasets import load_dataset

dataset = load_dataset("michaeldinzinger/webfaq-v2", "eng")["default"]
print(dataset[0])

Fields

• id: Unique identifier
• origin: Website origin (scheme + host)
• url: Source URL
• question: Natural-language question
• answer: Corresponding answer
• title: Webpage title (new in v2.0)
• description: Webpage description (new in v2.0)
• qa_similarity_score: Semantic similarity between question and answer
• topic (optional): Topic label
• question_type (optional): Question category

Note: No official train/validation/test split is provided for the raw QA dataset.

---

Bilingual Alignments

The bitext dataset is released separately:

👉 https://huggingface.co/datasets/michaeldinzinger/webfaq-v2-bitexts

Alignment method:

• LaBSE embeddings
• Nearest-neighbor retrieval
• Similarity threshold ≥ 0.9

Results:

• 14.3M aligned QA pairs
• Strong growth in non-English language pairs
• Includes underrepresented combinations (e.g., Marathi–Telugu)

---

Hard Negatives Dataset

Released separately:

👉 https://huggingface.co/datasets/IrvinTopi/WebFAQHardNegatives

Mining Pipeline

Two-stage approach :

1. BM25 retrieval (top 200 candidates)
2. Cross-encoder reranking using BGE-m3

Each instance contains:

(query, positive, pos_score, negatives[], neg_scores[])

Total size:

• 1.25M training samples
• 20 languages
• 200 negatives per query

---

Training Strategies Enabled

WebFAQ 2.0 supports two main dense retrieval training paradigms:

🔹 Contrastive Learning

• MultipleNegativesRankingLoss (MNR)
• Random negatives
• Hard negatives (Top-k or denoised)

🔹 Knowledge Distillation

• MarginMSE loss
• Uses cross-encoder scores directly
• Strong improvements in non-English retrieval

Experiments show:

• Random negatives remain surprisingly strong in contrastive setups.
• Knowledge distillation improves multilingual robustness but may slightly reduce English performance .

---

Examples

{
  "id": "example_id",
  "origin": "https://example.com",
  "url": "https://example.com/faq",
  "question": "How old do I have to be to rent a car in Girona?",
  "answer": "You must be at least 21 years old to rent a car in Girona.",
  "title": "Car Rental FAQ – Girona",
  "description": "Frequently asked questions about renting a car in Girona.",
  "qa_similarity_score": 0.93,
  "topic": "Traveling and Hospitality",
  "question_type": "Experience"
}

---

Considerations

• ⚠️ No guarantee of factual correctness
• ⚠️ Contains duplicates or near-duplicates
• ⚠️ Hard negatives may contain false negatives
• ⚠️ Language detection may be imperfect
• 📜 Derived from public web pages – respect original website terms

---

License

Open Data Commons Attribution License (ODC-By)

---

Citation

If you use WebFAQ 2.0, please cite:

@inproceedings{dinzinger2025webfaqv2,
  title={WebFAQ 2.0: A Multilingual QA Dataset with Mined Hard Negatives for Dense Retrieval},
  author={Dinzinger, Michael and Caspari, Laura and Salman, Ali and Topi, Irvin and Mitrović, Jelena and Granitzer, Michael},
  year={2025}
}

You may also cite the original WebFAQ paper:

@misc{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 Mitrović and Michael Granitzer},
    year={2025},
    eprint={2502.20936},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

---

Contact

For inquiries:

• michael.dinzinger@uni-passau.de
• HuggingFace Discussions
• GitHub: https://github.com/padas-lab-de/webfaq

---

WebFAQ 2.0 is part of a long-term effort aligned with the Open Web Index initiative and will continue to grow as new structured FAQ data becomes available .

Happy researching! 🚀