davanstrien/fineweb-curated-1m

FineWeb, 1M pages curated with a small decision model

The first 1,020,000 pages of one FineWeb crawl (CC-MAIN-2025-26), each with four curation answers from davanstrien/ultrabert-380m-curation: topic, format, educational value and reasoning depth. It is a worked example of curating a web dataset on Hugging Face Jobs: six GPU Jobs scored every page in about 19 minutes, for about $5.

Config	Rows	What it is
scored (default)	1,020,000	every page with its answers: look at these and choose your own filter
subset	63,788 (6.3 %)	one example filter: educational value "useful for education" or higher, and not spam, a content listing, a product page or incomplete

from datasets import load_dataset
scored = load_dataset("davanstrien/fineweb-curated-1m", "scored", split="train", streaming=True)
# your own filter, e.g. tutorials about science with at least some educational value
mine = scored.filter(lambda r: r["format"] == "Tutorial / How-To Guide"
                     and r["topic"] == "Science, Math & Technology" and r["edu_level"] >= 2)
subset = load_dataset("davanstrien/fineweb-curated-1m", "subset", split="train")   # the example filter

Columns

Column	Meaning
text, id, dump, url, date, file_path, language, language_score, token_count	from FineWeb, unchanged
topic, topic_p	most likely of 24 WebOrganizer topics, and its probability
format, format_p	most likely of 24 WebOrganizer formats, and its probability
edu, edu_p	most likely educational-value level (6 levels, "not educational" to "outstanding educational value"), and its probability
edu_level, edu_expected	that level as 0–5, and the expected level (probability-weighted mean)
reasoning, reasoning_p	most likely reasoning depth (Essential-Web scale), and its probability
cut_at_3000_chars	true when the page is longer than 3,000 characters; the model saw only the first 3,000 (the training length for web text). text is always the full page
in_subset	the subset rule: edu_level >= 3 and format not in Spam / Ads, Content Listing, Product Page, Incomplete Content

Only the top answer per question is kept, not the full distribution. To filter on other options or thresholds, rescore with the model.

What the answers look like

Share of the 1,020,000 scored pages (top answer per question):

Educational value	Pages	Format (top 5)	Pages	Topic (top 5)	Pages
not educational	16.1 %	Product Page	25.1 %	Finance & Business	10.9 %
minimal educational value	57.0 %	News Article	11.0 %	Home & Hobbies	9.5 %
some educational value	20.1 %	Organizational Announcement	8.1 %	Health	7.5 %
useful for education	6.3 %	Tutorial / How-To Guide	7.8 %	Education & Jobs	7.1 %
highly educational	0.5 %	Knowledge Article	7.6 %	Fashion & Beauty	5.2 %
outstanding educational value	1 page

Reasoning depth: Basic 57.9 %, No reasoning 22.1 %, Intermediate 15.5 %, Advanced 1.2 %, Indeterminate 3.3 %. 36.1 % of pages are longer than 3,000 characters, so the model saw only their beginning.

The model almost never picks the top educational level, and a quarter of the pages are product pages. The subset rule (level 3 or higher, not a product page or similar) keeps 6.3 %.

How it was made

OUT="hf://buckets/YOUR_NAME/YOUR_BUCKET/fineweb-1m"
uv run --python 3.12 --with "datatrove[io,processing]==0.10.1" --with "huggingface_hub>=1.32" \
    datatrove_curation.py --out "$OUT" --source hf://datasets/HuggingFaceFW/fineweb/data/CC-MAIN-2025-26 \
    --glob "000_0000*.parquet" --text-key text --id-key id --limit 170000 --tasks 6 --timeout 40m

datatrove_curation.py is in the model repo. It reads the first 10 FineWeb files with datatrove, 170,000 pages per task over six RTX PRO 6000 Jobs, scores each page with the model's packed GPU path, and writes the scored pages and the subset to a storage bucket. The flat columns here come from datatrove's metadata field.

Limitations

• The answers are a small model's guesses. On held-out pages it agrees with the WebOrganizer labels on 84 % of topics and 74 % of formats, and it ranks FineWeb-Edu educational value with a Spearman correlation of 0.78 (see the model card). Treat the answers as sorting signals.
• The subset rule is a rule of thumb. This release does not show that the subset makes better training data than FineWeb or FineWeb-Edu.
• The model saw only the first 3,000 characters of each page.
• Pages are the first files of one crawl, not a random sample of FineWeb.

Licence

FineWeb is released under ODC-By 1.0 and is subject to the Common Crawl terms of use; the same applies here. The added answer columns are released under ODC-By 1.0 too.