README.md

metadata

license: cc-by-nc-4.0
language:
  - en
pretty_name: FineQuestions
size_categories:
  - 1B<n<10B
task_categories:
  - text-retrieval
  - text-classification
  - question-answering
tags:
  - questions
  - intent
  - web-as-corpus
  - commoncrawl
  - fineweb
source_datasets:
  - HuggingFaceFW/fineweb
extra_gated_heading: Access FineQuestions
extra_gated_description: >-
  FineQuestions is released for non-commercial use under CC BY-NC 4.0. Please
  read the terms below before downloading.
extra_gated_prompt: >-
  FineQuestions is derived from FineWeb (ODC-By 1.0), which is built from Common
  Crawl. By requesting access you agree to: (1) use FineQuestions only for
  non-commercial purposes under CC BY-NC 4.0, with attribution to FineQuestions
  and FineWeb; (2) follow the FineWeb license and the Common Crawl Terms of Use
  for any page content you retrieve through the released document IDs and URLs;
  (3) not attempt to identify or contact individuals from the data; and (4)
  acknowledge that the questions are unfiltered web text that may contain
  profanity, offensive or explicit content, and personal names.
extra_gated_fields:
  I agree to use FineQuestions for non-commercial purposes only: checkbox
  I agree to follow the FineWeb license and Common Crawl Terms of Use for any source content I retrieve: checkbox
  I understand the data is unfiltered web text and may contain offensive or sensitive content: checkbox
extra_gated_button_content: Agree and access
configs:
  - config_name: sample
    default: true
    data_files:
      - split: train
        path: sample/sample.parquet
  - config_name: count_ge50
    data_files:
      - split: train
        path: data/gte50/*.parquet
  - config_name: count_10_49
    data_files:
      - split: train
        path: data/10_49/*.parquet
  - config_name: count_5_9
    data_files:
      - split: train
        path: data/5_9/*.parquet
  - config_name: count_3_4
    data_files:
      - split: train
        path: data/3_4/*.parquet
  - config_name: count_1_2
    data_files:
      - split: train
        path: data/1_2/*.parquet

FineQuestions

1.49 billion unique questions mined from the web, each linked to its source documents in FineWeb: a map of what people actually ask.

If Common Crawl is a memory of the web, FineQuestions is a map of its intentions.

FineQuestions extracts every question from FineWeb (110 CC-MAIN snapshots, 2013–2025), then normalizes and aggregates them. Each question comes with its occurrence count, source pointers and, for questions seen at least 5 times, content categories.

Stage	Count
FineWeb source documents	25.9B
Raw extracted questions	13,396,141,075
Unique normalized questions	1,493,938,805
Occurring ≥5 times, with content categories	599,720,275

Configs

The five count tiers don't overlap and together make up the full dataset. sample is the default.

Config	Occurrence count	Rows	Content categories
sample (default)	all tiers, 20K each	100,000	≥5 only
count_ge50	≥50	27,375,443	yes
count_10_49	10–49	302,933,413	yes
count_5_9	5–9	269,411,419	yes
count_3_4	3–4	238,021,805	null
count_1_2	1–2	656,196,725	null

from datasets import load_dataset
# quick look: 100K rows, 20K random from each tier, one ~100MB file
sample = load_dataset("BodhiumLabs/FineQuestions", "sample", split="train")
# one tier: stream rather than download
ds = load_dataset("BodhiumLabs/FineQuestions", "count_ge50", split="train", streaming=True)
# several tiers at once, e.g. every question seen at least 5 times (599.7M rows, all with categories);
# use "data/*/*.parquet" for all 1.49B
ge5 = load_dataset("BodhiumLabs/FineQuestions", split="train", streaming=True,
                   data_files=["data/gte50/*.parquet", "data/10_49/*.parquet", "data/5_9/*.parquet"])

Schema

Column	Type	Notes
norm_qn_id	string	xxhash64 of the normalized question (key)
question	string	Normalized question text
count	int64	Total occurrences across all 110 snapshots (a page recrawled in several snapshots counts once per snapshot; a question repeated on one page counts each time)
tier	string	gte50 / 10_49 / 5_9 / 3_4 / 1_2
samples	list⟨struct⟩	The 100 most recent occurrences by crawl date: uuid, url, host, date, question (as written on the page, before normalization), word_count, content_length
content_cat1..5	string	Top-5 content categories (26-category taxonomy). Null when count < 5.
content_conf1..5	double	Classifier confidence for each category. Null when count < 5.

Files are data/<tier>/<shard>.parquet. The 256 shards come from hash-partitioning norm_qn_id, so a given question is always in the same shard number across tiers.

Page content: pointers, not pages

We do not redistribute page text. Each sample's uuid is a FineWeb document ID, so content can be rejoined exactly:

-- DuckDB
SELECT q.question, s.url, fw.text
FROM 'hf://datasets/BodhiumLabs/FineQuestions/data/gte50/00.parquet' q, UNNEST(q.samples) AS t(s)
JOIN 'hf://datasets/HuggingFaceFW/fineweb/data/<CC-MAIN-dump>/*.parquet' fw
  ON fw.id = '<urn:uuid:' || s.uuid || '>'
LIMIT 100;

How it was built

Regex question extraction (sentences ending in ?, 5–100 words, starting with a question word or containing a question keyword; 0.59 questions per document) over FineWeb → normalization (lowercase, punctuation removed, whitespace collapsed) → xxhash64 → partition-then-aggregate map-reduce (256 buckets) keeping each question's total count and 100 most recent occurrences → GLiNER2 zero-shot content classification for questions occurring ≥5 times.

Known limitations (read before using)

• Frequency ≠ importance. High-count questions skew toward SEO and boilerplate text, so naive top-N selection surfaces junk. count measures how often a question is written on crawled pages, not how often it is asked.
• Extraction quality: on a human-validated sample of the count ≥5 slice, 79% of rows are genuine questions. The count <5 tiers contain more noise.
• Content categories are zero-shot labels, not ground truth. On a judged sample, content_cat1 was defensible for 70% of questions. Use them as filters.
• samples holds the 100 most recent occurrences, so for popular questions (max count 12.3M) it is a recent slice rather than a full list, and count > len(samples). Rejoin through FineWeb for full page text.
• Deduplication is exact-match on normalized text, so paraphrases remain separate rows.
• English only. The data inherits FineWeb's crawl and filtering biases.

Content warning

We did not filter questions for content beyond the filtering FineWeb applies. The data includes profanity, offensive and explicit questions, spam, and questions on sensitive subjects (health, religion, politics) at roughly the rates they appear on the web. Filter for your own use case before training or evaluating models.

Personal information

All text is excerpted from FineWeb, which replaced email addresses and public IP addresses with placeholders. A structured-PII scan of FineQuestions found emails, phone numbers and ID-like strings in under 0.05% of rows, mostly false positives such as product codes. Personal names are not redacted, consistent with the source corpus. Questions may contain sensitive content that appeared on public web pages.

Removal requests: email vincent@bodhiumlabs.com with the norm_qn_id or URL.

License

FineQuestions is released under CC BY-NC 4.0: you may use, share and adapt it for non-commercial purposes, with attribution. Commercial use is not permitted.

FineQuestions is derived from FineWeb, which is released under ODC-By 1.0. Page text rejoined from FineWeb remains under FineWeb's terms, and use is also subject to the Common Crawl Terms of Use. Please credit both FineQuestions and FineWeb.

Citation

A technical report describing the pipeline and experiments is coming soon on arXiv. Until then, please cite this dataset page:

@misc{bodhiumlabs2026finequestions,
  title  = {FineQuestions: 1.49 Billion Questions from the Web},
  author = {McCloskey, Vincent and Srinivasan, Krishna},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/BodhiumLabs/FineQuestions}}
}