wayu-ai/thai-commoncrawl-index

Thai Common Crawl Index (2019–2026)

An index of every page Common Crawl detected as Thai across 70 monthly crawls, from January 2019 (CC-MAIN-2019-04) to August 2026 (CC-MAIN-2026-30).

932,874,727 page captures · 450,971,497 unique URLs · 6,997,185 hosts · 6,674,969 domains

Each row records where the page lives inside Common Crawl's WARC archives — file name, byte offset, and record length — so you can fetch exactly the pages you want with HTTP range requests, without scanning or downloading the archives themselves.

This is an index, not a text corpus

The dataset contains URLs and byte pointers only. There is no page text or HTML here. It answers "which Thai pages exist in Common Crawl, and where are they?" — you then fetch the subset you care about (see Fetching page content).

That separation is the point. Common Crawl's own columnar index is hundreds of gigabytes per crawl and covers every language; isolating Thai across 70 crawls means scanning several terabytes. This dataset is the result of that scan, distributed as ~31 GB of Parquet, so the expensive step happens once instead of once per user.

Typical uses:

• Building Thai web corpora — filter to the domains and crawls you want, then fetch only those records.
• Studying the Thai web — domain composition, per-site scale, and how both shift across an eight-year span.
• Targeted retrieval — locate historical captures of specific sites or URLs.

Dataset structure

One row per Thai page capture. Parquet, partitioned by crawl under data/crawl=CC-MAIN-YYYY-WW/.

Column	Type	Description
host	string	Full hostname, e.g. news.thaipbs.or.th
registered_domain	string	Registrable domain (public-suffix aware), e.g. thaipbs.or.th
url	string	The page URL as crawled
fetch_time	timestamp	When Common Crawl fetched the page (UTC)
content_languages	string	CLD2 detection, comma-separated, up to 3 codes, most confident first — e.g. tha, tha,eng, eng,tha
warc_filename	string	Path to the WARC file within Common Crawl, e.g. crawl-data/CC-MAIN-2026-30/segments/…/CC-MAIN-….warc.gz
warc_record_offset	int32	Byte offset of this page's record in that file
warc_record_length	int32	Record length in bytes

The crawl id is the partition directory name. It is also embedded in warc_filename, so it is recoverable even when a loader drops the partition column (see Loading).

Domain-level aggregates are not shipped as a separate table — group by host or registered_domain to build whichever rollup you need.

Loading the data

DuckDB (recommended)

DuckDB reads the Hive partitioning, so you get a real crawl column and can push filters down to skip files entirely.

SELECT host, url, warc_filename, warc_record_offset, warc_record_length
FROM read_parquet('data/crawl=*/*.parquet', hive_partitioning = true)
WHERE crawl = 'CC-MAIN-2026-30'
  AND registered_domain = 'thaipbs.or.th'
  AND starts_with(content_languages, 'tha');

Straight from the Hub, without downloading everything first — here, the biggest Thai domains in one crawl:

import duckdb
con = duckdb.connect()
con.execute("INSTALL httpfs; LOAD httpfs;")
con.execute("""
    SELECT registered_domain,
           count(*) AS captures,
           count(DISTINCT url) AS uniq_urls
    FROM read_parquet(
      'hf://datasets/wayu-ai/thai-commoncrawl-index/data/crawl=CC-MAIN-2026-30/*.parquet')
    WHERE starts_with(content_languages, 'tha')
    GROUP BY 1 ORDER BY captures DESC LIMIT 50
""").df()

datasets

from datasets import load_dataset

# Stream rather than materializing ~933M rows
ds = load_dataset("wayu-ai/thai-commoncrawl-index", split="train", streaming=True)
print(next(iter(ds)))

datasets does not turn Hive partition directories into a column, so crawl will be absent when loading this way. Recover it from the path segment in warc_filename:

row["warc_filename"].split("/")[1]   # 'CC-MAIN-2026-30'

To pull only certain crawls, fetch just those files:

from huggingface_hub import snapshot_download

snapshot_download("wayu-ai/thai-commoncrawl-index", repo_type="dataset",
                  allow_patterns=["data/crawl=CC-MAIN-2026-*/*.parquet"])

Fetching page content

Common Crawl serves WARC files over HTTP with range support, and each record is an independently gzipped member — so a range request returns a standalone, decompressible record.

import io, requests
from warcio.archiveiterator import ArchiveIterator

def fetch_page(row):
    """Return the raw HTTP response bytes for one pointer row."""
    start = row["warc_record_offset"]
    end = start + row["warc_record_length"] - 1
    resp = requests.get(
        "https://data.commoncrawl.org/" + row["warc_filename"],
        headers={"Range": f"bytes={start}-{end}"},
        timeout=60,
    )
    resp.raise_for_status()
    for record in ArchiveIterator(io.BytesIO(resp.content)):
        return record.content_stream().read()

html = fetch_page(row)

Practical notes:

• Be gentle with data.commoncrawl.org. It sits behind a WAF that throttles aggressive clients with HTTP 403 (not 429). Keep concurrency modest and back off on 403/429/503.
• At scale, read from S3 instead. The same objects live at s3://commoncrawl/… under the AWS Open Data Program. From EC2 in us-east-1 there are no request or transfer charges, which is the only practical route for millions of records.
• Sort your work by warc_filename before fetching — records from the same file are far cheaper to retrieve together.

Coverage

Year	Crawls	Thai captures
2019	12	124,704,804
2020	9	105,160,815
2021	9	125,305,199
2022	6	92,496,089
2023	5	87,038,049
2024	10	141,882,673
2025	12	172,660,201
2026	7	83,626,897
Total	70	932,874,727

Yearly totals track how many crawls Common Crawl published that year, not the size of the Thai web. Per crawl, Thai captures stay within a narrow band across the whole span — 9,496,441 (CC-MAIN-2019-18) to 19,420,727 (CC-MAIN-2023-40).

Every crawl from CC-MAIN-2019-04 through CC-MAIN-2026-30 is included, with none skipped. The 2019 start is a hard boundary: content_languages was introduced in CC-MAIN-2018-34, and no per-page language detection exists in the index before it.

Domain distribution is steeply long-tailed. Of 6,674,969 domains, 59,308 have at least 1,000 unique URLs and 89,182 appear in at least half the crawls — those subsets are the realistic candidates for corpus building.

How it was built

A single query over Common Crawl's public columnar URL index (s3://commoncrawl/cc-index/table/cc-main/warc/), selecting subset = 'warc' rows whose content_languages contains tha, across the 70 crawls in range. The published rows are that query's output, unmodified apart from partitioning by crawl.

The result was cross-checked against an independent local scan of the same index built on a different engine; per-crawl page and host counts matched exactly on the crawls compared.

No page content was fetched, and no filtering, deduplication, or quality scoring was applied — this is a faithful index of what Common Crawl's language detector marked as Thai, and deliberately not a curated corpus.

Limitations

tha may be a secondary language on the page. A row is included if CLD2 lists Thai among its up-to-three detected languages, in any position. That is a deliberate recall-first choice, and it matters:

Selection	Captures	Share
Thai in any position (all rows)	932,874,727	100%
Thai-primary (content_languages starts with tha)	727,398,274	78.0%
Thai only (content_languages = 'tha')	181,037,965	19.4%

The most common value is tha,eng (502,066,753 rows). The 22% where Thai is not first include pages with only a fragment of Thai — a German-language insurance page tagged deu,eng,tha, for instance. For corpus work, filter on Thai-primary, and consider re-running language identification on the extracted text.

Language detection is imperfect. CLD2 is fast but errs on short pages, boilerplate- heavy pages, and mixed-script content. Treat content_languages as a strong prior, not ground truth.

Thai-language does not mean Thai-authored. Some of the highest-volume "100% Thai" sites are machine-translated localizations of global platforms — the Thai subdomains of large travel, e-commerce, and listing services. They are genuinely Thai text and genuinely 100% Thai by ratio, but they read as translationese. No ratio or threshold separates them; if native-authored Thai is what you need, exclude such sites by name.

Captures are not unique pages. The same URL recrawled in 40 crawls yields 40 rows. 932,874,727 captures reduce to 450,971,497 unique URLs. Deduplicate on url — keeping the row with the latest fetch_time — before treating counts as page counts.

Page counts reflect crawl budget. How many pages a site has in Common Crawl is a function of Common Crawl's frontier and politeness settings, not the site's true size, traffic, or importance. Rankings here are rankings within Common Crawl.

Coverage of the Thai web is partial. Common Crawl honors robots.txt and misses login-walled content, most social platforms, and sites that block crawlers. A .th TLD filter would be a poor proxy for Thai content, and the reverse holds too: much Thai content sits on .com.

Pointers can go stale. They resolve against Common Crawl's archives as currently published. Nothing here guarantees the referenced page still exists on the live web.

Licensing and use

The index itself — URLs, hostnames, timestamps, and byte offsets — is factual metadata derived from Common Crawl's public index, released under ODC-BY 1.0. Please attribute both this dataset and Common Crawl.

The page content these pointers reference is not covered by that license. Retrieving and using it is governed by Common Crawl's Terms of Use and by the rights of the original publishers. If you build a corpus from it, you are responsible for copyright, privacy, and any applicable local law, and for honoring removal requests. Common Crawl provides a takedown process; content withdrawn there should be withdrawn from derivatives too.

Because this is a raw index of crawled web pages, it inevitably points at material that is low-quality, spam, adult, offensive, or factually wrong, and at pages containing personal information. Filtering is left to you, and is strongly recommended.

Citation

@misc{thai_commoncrawl_index,
  title        = {Thai Common Crawl Index (2019--2026)},
  author       = {wayu-ai},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/wayu-ai/thai-commoncrawl-index}}
}

Please also credit Common Crawl, whose archives and columnar index this is derived from:

@misc{commoncrawl,
  title        = {Common Crawl},
  howpublished = {\url{https://commoncrawl.org}}
}

Acknowledgements

Built entirely on Common Crawl's open archives and columnar index, and on the AWS Open Data Program that hosts them.