Boldt/fineweb2-de-dc

FineWeb2-DE Dense-Core — BOLDT Pre-Training Data

This dataset accompanies the paper Repetition over Diversity: High-Signal Data Filtering for Sample-Efficient German Language Modeling (EMNLP 2026). It is one of the BOLDT pre-training subsets, released together with the BOLDT models and the German LLM benchmarks.

Dense-Core is the most selective subset of the German split of FineWeb2. It is the intersection of all three quality tiers: it keeps only documents that are coherent, information-dense, and educational. It contains ~5% of FW2-DE and is the high-quality core that the BOLDT models are trained on for several epochs.

Dataset Summary

Statistic	Value
Source	German split of FineWeb2 (deu_Latn, ~496M documents)
Documents	24,446,297
Share of FW2-DE (%)	5.1
Tokens	28B
Mean document length (tokens)	1150 ± 2193

Token counts use the 32k-vocabulary BPE tokenizer trained on each subset in the paper.

Selection Criterion

Documents must pass all three filters at the same time:

Filter	Score field	Kept if (rounded score)
Coherence	coherence_score	≥ 3 (raw ≥ 2.5)
Information Value	iv_score	≥ 3 (raw ≥ 2.5)
Educational Quality	edu_score	≥ 3 (raw ≥ 2.5)

Scores come from the BOLDT quality classifiers (see below). Thresholds are applied to the classifier's predicted score rounded to the nearest integer.

How the subsets were built

The BOLDT data pipeline uses three hierarchical quality tiers to score all ~500M documents in FW2-DE. Ordered from least to most strict:

Tier	Targets	Score range	Dataset
Coherence	Structural and linguistic integrity, natural syntactic flow	1–3	Boldt/fineweb2-de-coherence
Information Value	Fact-bearing, content-rich, non-promotional text	1–4	Boldt/fineweb2-de-iv
Educational Quality	Teaching value and textbook-like clarity (FineWeb-Edu criteria)	0–5	Boldt/fineweb2-de-edu
Dense-Core	Intersection of all three tiers	–	Boldt/fineweb2-de-dc

The tiers can overlap: a document can belong to several subsets.

Annotation. The pipeline relies on the following recipe:

1. Llama-3.3-70B-Instruct annotated 500k randomly sampled FW2-DE documents, truncated to 4096 tokens.
   • Educational Quality uses the original FineWeb-Edu prompt.
   • Coherence and Information Value use a joint prompt defined in the paper (Appendix A, Figure 6).
2. Three regression heads on snowflake-arctic-embed-m-v2.0 were fine-tuned on these labels, one per tier. Inputs were truncated to 1024 tokens.
3. The classifiers then scored all ~500M FW2-DE documents. They are released here:

Classifier	Micro F1	Binary F1 at cutoff
Coherence (Boldt/fineweb2-de-coherence-classifier)	84.32	90.38
Information Value (Boldt/fineweb2-de-iv-classifier)	78.45	78.50
Educational Quality (Boldt/fineweb2-de-edu-classifier)	75.61	72.32

Statistics of all BOLDT subsets

Subset	Docs (M)	Yield (%)	Tokens	Doc length (tokens, µ ± σ)
FW2-DE (full pool)	496.0	100.0	–	–
Coherence (full / 100B sample)	300.6 / 138.3	60.6 / 27.8	– / 100B	730 ± 1540
Information Value	43.5	8.8	65B	1494 ± 2561
Educational Quality	30.2	6.1	33B	1087 ± 2103
Dense-Core	24.5	5.1	28B	1150 ± 2193

Models trained on this subset

• Boldt/Boldt-DC-350M: 350M, 200B tokens (7.2 epochs)
• Boldt/Boldt-DC-1B: 1B, 200B tokens (7.2 epochs)
• Boldt/Boldt-DC-350M-100BT: 350M, 100B tokens (3.6 epochs)
• Boldt/Boldt-DC-1B-100BT: 1B, 100B tokens (3.6 epochs)
• Boldt/Boldt-Phased-350M-100BT: 350M, 50B Random tokens followed by 50B Dense-Core tokens
• Boldt/Boldt-Phased-350M-200BT: 350M, 100B Random tokens followed by 100B Dense-Core tokens
• Boldt/Boldt-1B: 1B, Dense-Core plus ~6B tokens of German news from Fundus, 230B tokens, 4096 context

License

This dataset is derived from FineWeb2. Like FineWeb2, it is released under the ODC-By 1.0 license and is subject to Common Crawl's Terms of Use.

Citation

@inproceedings{aynetdinov-etal-2026-repetition,
    title = "Repetition over Diversity: High-Signal Data Filtering for Sample-Efficient {G}erman Language Modeling",
    author = "Aynetdinov, Ansar and Haller, Patrick and Akbik, Alan",
    editor = "Lee, Hung-Yi and Dev, Sunipa and Li, Jessy and Elliott, Desmond",
    booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
    month = oct,
    year = "2026",
    address = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.emnlp-main.1473/",
    doi = "10.18653/2026.emnlp-main.1473",
    pages = "33690--33710"
}

Please also cite FineWeb2, which this dataset is derived from:

@inproceedings{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 Kargaran, Amir Hossein and Raffel, Colin and Jaggi, Martin and Von Werra, Leandro and Wolf, Thomas},
  booktitle={Second Conference on Language Modeling},
  year={2025}
}