DataCraftsmanAustralia/fineweb-au-15BT

FineWeb AU 15BT

14.7 billion tokens of Australian web text in 25.2 million documents from 643,634 .au hosts. Every document comes from a host ending in .au, taken from FineWeb and cleaned for pretraining. The text covers 17 Common Crawl snapshots from CC-MAIN-2023-50 (December 2023) to CC-MAIN-2025-26 (June 2025), and each snapshot contributes the same number of tokens. The 15BT in the name is the token count, rounded.

Most open language models learn from web text in which Australian English is a small fraction. Models trained that way spell the American way and answer questions about superannuation, Medicare or state law with overseas assumptions. This dataset is a starting point for continued pretraining on Australian text. Data Craftsman, an Australian AI lab, built it to train Australis Ventus.

Documents	25,218,369
Tokens (GPT-2)	14,697,955,255
Hosts	643,634
Snapshots	17, each 0.865 billion tokens
Median document length	356 tokens
Files	3,289 parquet files, 30.1 GB
Licence	ODC-By 1.0

Usage

from datasets import load_dataset

# Stream the whole corpus
ds = load_dataset("DataCraftsmanAustralia/fineweb-au-15BT", split="train", streaming=True)
for row in ds.take(3):
    print(row["url"], row["token_count"])

# Load a single snapshot
ds = load_dataset("DataCraftsmanAustralia/fineweb-au-15BT", "CC-MAIN-2025-26", split="train")

Fields

Field	Type	Description
url	string	Source URL of the page
text	string	Page text as extracted by FineWeb
token_count	int64	FineWeb's GPT-2 token count for the text
dump	string	Common Crawl snapshot, such as CC-MAIN-2024-30
date	string	Crawl date and time, such as 2024-07-13T15:20:25Z

The fields and text are unchanged from FineWeb. This dataset selects rows. It does not rewrite them.

How it was built

Source. FineWeb is English web text from Common Crawl. It has already applied language identification, quality filters, MinHash deduplication within each snapshot, and anonymisation of email addresses and public IP addresses.

Extraction. We scanned 3,289 FineWeb shards from the 17 snapshots, holding 3.08 billion documents, and kept every row whose URL host ends in the label au. That keeps direct .au names as well as second level names such as .com.au and .gov.au. Three snapshots were scanned in full (CC-MAIN-2024-10, CC-MAIN-2024-18 and CC-MAIN-2024-22) and the rest in part. The raw extract held 80.4 million documents and 48.76 billion tokens, about 2.6% of the documents scanned.

Cleaning. Four steps turned the raw extract into this dataset.

1. Evaluation hold out removed. Every row whose URL is in the hold out list for our Australian dialect evaluation was removed: 2,687 URLs, matching 6,150 rows.
2. URL deduplication across snapshots. FineWeb deduplicates within each snapshot but not across them, so the same page can appear once per snapshot. Almost half the raw rows repeat a URL from another snapshot. One copy of each URL was kept, chosen at random.
3. Host cap. Each host was capped at 2,000 documents, chosen at random. In the raw extract, 4,761 hosts with more than 2,000 documents each held 32.3% of the text. The ABC alone had 205,984 documents.
4. Snapshot balancing. Each snapshot was capped at the token count of the smallest, 0.865 billion tokens, so no single snapshot of the web dominates.

All random choices used seed 7.

Stage	Documents	Tokens (billion)	Hosts
Raw .au extract	80,422,150	48.76	670,511
After removing evaluation documents	80,416,000	48.76	670,507
After URL deduplication	40,848,410	24.53	670,006
After the host cap	33,351,320	19.50	670,006
After snapshot balancing (this dataset)	25,218,369	14.70	643,634

Verification. A check run independently of the build found zero duplicate URLs and zero rows from the evaluation hold out. A contamination scan of every document against Terminal-Bench 2.1 and 4.0 found no canary string and no 13-gram overlap with any task. The scanner caught 10 of 10 planted test leaks before it was run on the corpus.

Composition

By domain type

Domain type	Documents	Share of documents	Tokens (billion)	Share of tokens	Hosts
.com.au (commercial)	21,227,124	84.2%	12.004	81.7%	559,690
.org.au (organisations)	1,591,989	6.3%	1.038	7.1%	33,401
.edu.au (education)	760,327	3.0%	0.483	3.3%	13,865
.gov.au (government)	734,221	2.9%	0.485	3.3%	7,558
.net.au	476,373	1.9%	0.401	2.7%	14,751
direct .au	293,625	1.2%	0.194	1.3%	11,797
.asn.au (associations)	93,789	0.4%	0.060	0.4%	1,644
.id.au (individuals)	24,136	0.1%	0.023	0.2%	729
other (.csiro.au, state community domains)	16,785	0.1%	0.010	0.1%	199

By document length

Length (tokens)	Share of documents	Share of tokens
under 150	17.9%	3.4%
150 to 499	45.0%	22.5%
500 to 1,499	30.8%	43.8%
1,500 to 4,999	5.9%	23.0%
5,000 and over	0.5%	7.3%

By host. No host has more than 1,806 documents. The top 10 hosts hold 0.07% of documents and the top 1,000 hold 6.23%. The median host contributes 7 documents, and 15.3% of hosts contribute a single document.

By snapshot

Snapshot	Documents	Tokens (billion)
CC-MAIN-2023-50	1,537,423	0.865
CC-MAIN-2024-10	1,447,434	0.865
CC-MAIN-2024-18	1,441,681	0.865
CC-MAIN-2024-22	1,417,934	0.865
CC-MAIN-2024-26	1,462,070	0.865
CC-MAIN-2024-30	1,479,818	0.865
CC-MAIN-2024-33	1,472,565	0.865
CC-MAIN-2024-38	1,481,483	0.865
CC-MAIN-2024-42	1,503,173	0.865
CC-MAIN-2024-46	1,488,681	0.865
CC-MAIN-2024-51	1,557,611	0.865
CC-MAIN-2025-05	1,502,672	0.865
CC-MAIN-2025-08	1,525,751	0.865
CC-MAIN-2025-13	1,513,452	0.865
CC-MAIN-2025-18	1,472,808	0.865
CC-MAIN-2025-21	1,476,594	0.865
CC-MAIN-2025-26	1,437,219	0.865

Limitations

• The .au domain is a proxy for Australian text. It includes content from outside Australia hosted on .au sites and misses Australian content on .com, .org and other generic domains.
• FineWeb's filters decide what was available. Its English identification and quality heuristics were not designed for Australian English, and we have not measured what they removed.
• Text quality is not scored. Beyond FineWeb's filters, deduplication and the host cap, no quality classifier was applied. More than four fifths of the text is from commercial sites.
• The host cap is flat. It treats a newsroom and a product catalogue the same way, so it removed much of the national news and university research text along with the low quality bulk.
• Some URLs are missing. Deduplication chose each URL's copy from a wider scan that also held four early 2023 snapshots outside this dataset. About 394,000 URLs whose chosen copy fell in those snapshots are not included, around 0.27 billion tokens.
• Token counts are GPT-2 counts. Newer tokenisers produce fewer tokens for the same text.

Personal and sensitive information

This is public web text. FineWeb anonymised email addresses and public IP addresses, and we did no further removal. Documents may still contain names, personal details, offensive content or factual errors. If you find content that should be removed, open a discussion on this repository with the URL.

Licence

The dataset keeps FineWeb's ODC-By 1.0 licence, since it is a filtered subset of FineWeb. Use is also subject to Common Crawl's terms of use. The underlying pages remain the work of their publishers.

Citation

@misc{jennings2026finewebau15bt,
  title        = {FineWeb AU 15BT: 14.7 billion tokens of Australian web text},
  author       = {Jennings, Michael},
  organization = {Data Craftsman},
  year         = {2026},
  url          = {https://huggingface.co/datasets/DataCraftsmanAustralia/fineweb-au-15BT}
}

Please also cite FineWeb:

@inproceedings{penedo2024fineweb,
  title     = {The {FineWeb} Datasets: Decanting the Web for the Finest Text Data at Scale},
  author    = {Penedo, Guilherme and Kydl{\'i}{\v{c}}ek, Hynek and Ben Allal, Loubna and Lozhkov, Anton and Mitchell, Margaret and Raffel, Colin and Von Werra, Leandro and Wolf, Thomas},
  booktitle = {Advances in Neural Information Processing Systems 37, Datasets and Benchmarks Track},
  year      = {2024}
}