FineWeb-CLaR-culture: culture-annotated multilingual web documents

29,695,269 documents across 5,954 locales (language–script–region combinations; 297 languages, 322 language–script pairs), each annotated with topic assignments and a cultural-taxonomy profile. Text derives from FineWeb and FineWeb-2 via the FineWeb-CLaR region-attribution pipeline.

Annotation pipeline

Per locale (up to 10,000 quality-filtered documents): BGE-M3 embeddings → FASTopic (500 topics, merged at cosine ≥ 0.85) → topic-to-taxonomy mapping (softmax over BGE-M3 similarity, τ = 0.05) onto the Cultural Topic Ontology of Liu, Gurevych & Korhonen (TACL 2025), with LLM adjudication of low-confidence topic labels (Qwen2.5-32B-Instruct, T=0; 99.97% of topics triggered) → per-document multi-label extraction (top-5 topics with probability ≥ 0.05).

The taxonomy: 3 L1 branches, 11 L2 cultural elements, 14 terminal leaves (the 10 childless L2 elements + 4 Values sub-types):

L1 branch	L2 elements
Ideational	concepts, knowledge, values (4 sub-types), norms & morals, artifacts
Linguistic	dialects, styles
Social	relationship, context, communicative goals, demographics

Layout

Hive-partitioned parquet (zstd), one file per locale:

data/language_script=deu_Latn/region=MD/deu_Latn-MD.parquet

Columns

column	type	description
doc_idx	int32	document index within the locale
locale	str	<lang>_<Script>-<REGION>, e.g. deu_Latn-MD
language, language_code, script, family	str	language metadata
region_name, region_code	str	region metadata (code also in the partition key)
text	str	document text (200–50,000 chars, lang-confidence ≥ 0.8)
topic_labels	list[str]	top words of each assigned topic (≤ 5 topics/doc)
topic_probs	list[float]	probability of each assigned topic
cto_primary	str	taxonomy leaf of the top topic (softmax projection, 14 leaves)
cto_adjudicated	str	LLM-adjudicated taxonomy leaf of the top topic (Qwen2.5-32B-Instruct, T=0; falls back to the softmax label for the 0.03% of topics never adjudicated)
cto_l1_distribution	str (JSON)	document mass over the 3 L1 branches
cto_l2_distribution	str (JSON)	document mass over the 11 L2 elements
cto_leaf	str	argmax of the L2 distribution
cto_leaf_confidence	float	probability mass of cto_leaf
cto_leaf_selected	str or null	cto_leaf if confidence ≥ 0.30, else null

On cto_leaf_confidence: the soft topic→taxonomy mapping keeps document distributions flat — max-leaf mass ranges ≈ 0.20–0.35 (uniform baseline 1/11 ≈ 0.09) and essentially never exceeds 0.4. The prebuilt cto_leaf_selected uses a 0.30 cutoff (≳ 3× uniform; keeps roughly the top third of documents corpus-wide, with strong per-locale variance). For another operating point, re-threshold on cto_leaf_confidence.

Usage

Stream everything

from datasets import load_dataset
ds = load_dataset("Yusser/FineWeb-CLaR-culture", split="train", streaming=True)

Efficiently extract text (partition pruning + column projection + predicate pushdown)

Only the requested partitions, columns, and row groups are downloaded:

import pyarrow.dataset as pad

d = pad.dataset("hf://datasets/Yusser/FineWeb-CLaR-culture/data", partitioning="hive")

# all German(y)-script docs from Moldova, text + selected leaf only
tbl = d.to_table(
    columns=["text", "cto_leaf_selected", "cto_leaf_confidence"],
    filter=(pad.field("language_script") == "deu_Latn") & (pad.field("region") == "MD"),
)
texts = tbl.column("text").to_pylist()

# confidently "values"-themed docs across ALL locales, text only
values_docs = d.to_table(
    columns=["text", "locale"],
    filter=(pad.field("cto_leaf") == "L2_values") & (pad.field("cto_leaf_confidence") >= 0.30),
)

Download one language (or locale) locally

from huggingface_hub import snapshot_download
snapshot_download(
    "Yusser/FineWeb-CLaR-culture", repo_type="dataset", local_dir="fineweb_clar",
    allow_patterns=["data/language_script=jpn_Jpan/*"],   # or ".../region=EG/*"
)

References

@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 = {NeurIPS Datasets and Benchmarks},
  year      = {2024}
}

@article{penedo2025fineweb2,
  title   = {FineWeb2: One Pipeline to Scale Them All -- Adapting Pre-Training Data Processing to Every Language},
  author  = {Penedo, Guilherme and Kydl{\'i}{\v{c}}ek, Hynek and Sabol{\v{c}}ec, Vinko and Messmer, Bettina and Foroutan, Negar and Jaggi, Martin and von Werra, Leandro and Wolf, Thomas},
  journal = {arXiv preprint arXiv:2506.20920},
  year    = {2025}
}

@inproceedings{wu2024fastopic,
  title     = {FASTopic: Pretrained Transformer is a Fast, Adaptive, Stable, and Transferable Topic Model},
  author    = {Wu, Xiaobao and Nguyen, Thong and Zhang, Delvin Ce and Wang, William Yang and Luu, Anh Tuan},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2024}
}

@article{liu2025culturally,
  title   = {Culturally Aware and Adapted {NLP}: A Taxonomy and a Survey of the State of the Art},
  author  = {Liu, Chen Cecilia and Gurevych, Iryna and Korhonen, Anna},
  journal = {Transactions of the Association for Computational Linguistics},
  volume  = {13},
  pages   = {652--689},
  year    = {2025}
}

@article{chen2024bge,
  title   = {BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation},
  author  = {Chen, Jianlv and Xiao, Shitao and Zhang, Peitian and Luo, Kun and Lian, Defu and Liu, Zheng},
  journal = {arXiv preprint arXiv:2402.03216},
  year    = {2024}
}

Provenance & license

Text originates from FineWeb / FineWeb-2 (ODC-By 1.0, subject to the CommonCrawl ToU); this dataset redistributes it with machine-generated annotations under the same ODC-By 1.0 terms. Annotations are fully automatic — no human labeling.