FineWeb-C: Educational content in many languages, labelled by the community

Multilingual data is better together!

Note: This datasets and the dataset card are works in progress. You can help contribute to the dataset here and join the community discussions in rocket chat!

What is this?

This is a collaborative, community-driven project that expands upon the FineWeb2 dataset. Our goal is to create high-quality educational content annotations across hundreds of languages.

By enhancing web content with these annotations, we aim to improve the development of Large Language Models (LLMs) in all languages, making AI technology more accessible and effective globally.

The annotations in this dataset will help train AI systems to automatically identify high-quality educational content in more languages and in turn help build better Large Language Models for all languages.

What the community is doing:

• For a given language, look at a page of web content from the FineWeb2 dataset in Argilla.
• Rate how educational the content is.
• Flag problematic content i.e. content that is malformed or in the wrong language.

Once a language reaches 1,000 annotations, the dataset will be included in this dataset! Alongside rating the educational quality of the content, different language communities are discussing other ways to improve the quality of data for their language in our rocket chat discussion channel.

What's been done so far?

So far 318 members of the Hugging Face community have submitted 32,863 annotations.

The following languages have reached the 1,000 annotation threshold to be included in the dataset. We'll keep updating this dataset as more annotations are added!

Language Code	Language Name	Completed Annotations	Annotators
arb_Arab	Standard Arabic	1000	10
ary_Arab	Moroccan Arabic	1000	15
arz_Arab	Egyptian Arabic	1000	9
bar_Latn	Bavarian	1000	1
cmn_Hani	Mandarin Chinese	1000	3
dan_Latn	Danish	1000	18
fas_Arab	Persian	1000	3
gmh_Latn	Middle High German	1000	1
hin_Deva	Hindi	1000	3
lvs_Latn	Standard Latvian	1000	5
rus_Cyrl	Russian	1000	4
tat_Cyrl	Tatar	1000	7

You can help contribute to the dataset here.

Below is an overview of the number of annotations submitted for each language (updated daily).

Why are we doing this?

There are many languages in the world where no high quality LLMs exist. Having high quality data is a central part of building high quality LLMs. FineWeb2 is a crucial step in improving the availability of high quality data for many languages. We plan to go a step further.

Fineweb-Edu for every language?

FineWeb-Edu is a dataset built on the original FineWeb dataset. The dataset was constructed by developing an educational quality classifier using annotations generated by LLama3-70B-Instruct and using this classifier to retain only the most educational web pages.

FineWeb-Edu outperforms FineWeb on popular benchmark. Crucially, using this approach reduces the amount of data needed to train a high quality LLM reducing the barrier to building a high quality LLM for many languages.

We want to make it possible to build FineWeb-Edu datasets for all the worlds languages. To do this we need annotations in order to train an educational quality classifier.

This in turn will allow us to build the next generation of Large Language Models for many languages.

Why not use LLMs to annotate the data?

For high resources languages, using an LLM to generate educational quality annotations can be a good solution. However, for many languages LLMs are not able to generate high quality annotations — or we don't have enough data to validate whether the annotations are correct.

How can I help?

You can help by contributing to the dataset here and join the community discussions in rocket chat!

Why would I bother to contribute to this dataset?

Your contributions directly shape the future of AI in your language. Here's why this matters:

1. Break the AI language barrier: Most commercial AI companies focus on profitable languages, leaving many communities behind. Your work helps bring AI capabilities to more languages.

2. Keep it open: Unlike proprietary datasets locked away by companies, FineWeb2-C is an open dataset. This means anyone can use it to build AI systems that truly serve their community's needs. Through this open approach we also learn about which approaches work best for different languages.

3. Be part of something bigger: Just as Wikipedia showed how volunteers can build invaluable resources, the Hugging Face community has created numerous open models and datasets. You're joining a movement to democratize AI technology.

Every annotation counts. Whether you can contribute ten minutes or ten hours, your input helps build a more inclusive future for AI technology 🤗

Who contributed to this dataset so far?

These are the top 10 contributors to this release of the dataset. Make sure to give them a follow on the Hub to show your appreciation!

Hugging Face Username	Submissions
stefan-it	2,011
hasnachouikhi	1,865
catastropiyush	1,053
vikkormallansohn	1,000
rasgaard	1,000
Maani	985
paperplanedeemo	978
JakobBlaa	978
anhha9	927
Aivis	894

Data work is the under appreciated foundation of AI and ML. This dataset is built by the community for the community. Below is a leaderboard that is updated daily and shows all the contributors to this annotation effort.

Language-specific Contributors

Below you can find a list of all the contributors to this release of the dataset for each language ❤️

Detailed Contributor Statistics for each language

Bavarian (bar_Latn)

User Statistics Table (Minimum 1 submissions)

Username	Submissions
stefan-it	1000

Danish (dan_Latn)

User Statistics Table (Minimum 1 submissions)

Username	Submissions
rasgaard	1000
JakobBlaa	978
saattrupdan	200
FrLars21	80
markhougaard	72
KennethEnevoldsen	44
Apasalic	33
tqvist	33
cnila	31
Soeren-B	28
KristianL	22
mathiasn1	16
ITK-dev	12
jannikskytt	8
AndreasLH	7
perlausten	5
sorenmulli	3
organicoder	1

Egyptian Arabic (arz_Arab)

User Statistics Table (Minimum 1 submissions)

Username	Submissions
mmhamdy	734
aishahamdy	141
oumayma03	54
omarelshehy	46
ghada00	14
heba1998	10
chemouda	3
aammari	2
amreleraqi	1

Hindi (hin_Deva)

User Statistics Table (Minimum 1 submissions)

Username	Submissions
catastropiyush	926
pp	73
Urmish	1

Mandarin Chinese (cmn_Hani)

User Statistics Table (Minimum 1 submissions)

Username	Submissions
paperplanedeemo	978
guokan-shang	12
AdinaY	10

Middle High German (gmh_Latn)

User Statistics Table (Minimum 1 submissions)

Username	Submissions
stefan-it	1000

Moroccan Arabic (ary_Arab)

User Statistics Table (Minimum 1 submissions)

Username	Submissions
Ihssane123	499
imomayiz	234
NouhailaChab05	120
nouamanetazi	58
master12gx	37
oumayma03	21
Overowser	14
SoufianeDahimi	12
adnananouzla	11
alielfilali01	3
staghado	3
olafdil	2
maghwa	2
0xTechVio	1
maggierphunt	1

Persian (fas_Arab)

User Statistics Table (Minimum 1 submissions)

Username	Submissions
Maani	985
mehrdadazizi	14
kargaranamir	1

Russian (rus_Cyrl)

User Statistics Table (Minimum 1 submissions)

Username	Submissions
kitano-o	593
kristaller486	396
knyazer	9
alialek	5

Standard Arabic (arb_Arab)

User Statistics Table (Minimum 1 submissions)

Username	Submissions
hasnachouikhi	1000
alielfilali01	4

Standard Arabic (arb_Arab)

User Statistics Table (Minimum 1 submissions)

Username	Submissions
hasnachouikhi	865
chemouda	102
oumayma03	12
ahmedselhady	9
staghado	7
alielfilali01	4
YassineL	2
maggierphunt	1

Standard Latvian (lvs_Latn)

User Statistics Table (Minimum 1 submissions)

Username	Submissions
Aivis	894
slckl	48
finnayeet	33
zemais	26
minem99	2

Tatar (tat_Cyrl)

User Statistics Table (Minimum 1 submissions)

Username	Submissions
tagay1n	515
gaydmi	313
inov8	126
iamdweebish	42
Giniyatullina	6
Empirenull	3
Khusaenov	1

Using this dataset

The dataset has a default config that contains all the language and configs per language.

To download the dataset using the Hugging Face datasets library, you can use the following code:

from datasets import load_dataset

dataset = load_dataset("data-is-better-together/fineweb-c-edu")

To download a specific language, you can use the following code:

dataset = load_dataset("data-is-better-together/fineweb-c-edu", language="cmn_Hani")

You can also download the dataset using Pandas

import pandas as pd

# Login using e.g. `huggingface-cli login` to access this dataset
df = pd.read_parquet("hf://datasets/data-is-better-together/fineweb-c-edu/arb_Arab/train-00000-of-00001.parquet")

or polars

import polars as pl

# Login using e.g. `huggingface-cli login` to access this dataset
df = pl.read_parquet('hf://datasets/davanstrien/fineweb-c-exported-data-test/arb_Arab/train-00000-of-00001.parquet')

Data Fields

The dataset contains the following columns:

Column Name	Type	Description
id	string	A unique identifier for each annotation record
text	string	The text of the web page
educational_value_labels	list[string]	A list of labels indicating the educational value of the web page rated by the community
annotator_ids	string	A string ID for the annotator
problematic_content_label_present	boolean	A flag indicating the presence of at leaste one 'problematic' label being assigned to the text
problematic_content_label_agreement	float	The agreement of the annotator with the problematic content label
language_names	str	The name of the language page
language_code	str	The code of the language

The main things to note (we'll update this as we get more data)

• Some languages already have multiple annotations per page. So far we haven't done any processing on these rows so people are free to calculate the agreement of the annotators in whatever way they want.
• For languages with many active annotators, we may increase the overlap of annotations over time to further improve the quality of the dataset.
• Some languages contain many problematic content labels. These often occur when the language detection was not correct. There is a problematic_content_label_present boolean column that indicates if the page contains at least one problematic content label. If you want to remove these rows you can do so by filtering on this column. Alternatively, you can use the problematic_content_label_agreement column to filter on the agreement of the annotators i.e. only remove rows where the annotators agree on the problematic content label. For many of the most active language efforts we're working with the community to improve the quality of the data so we hope the number of problematic content labels will decrease over time.

Licensing Information

The dataset is released under the Open Data Commons Attribution License (ODC-By) v1.0 license. The use of this dataset is also subject to CommonCrawl's Terms of Use.

Citation

Citation information needs to be added

Last Updated

2024-12-20