danish-foundation-models/propella-bekko-data

Propella-Bekko Data

The data behind the propella-bekko document annotators: propella-1-4b annotations for 844,912 documents from the Dynaword corpora, FineWeb-2 and FinePDFs, published together with the documents themselves.

Two held-out splits annotate the same 1,000 documents independently of the teacher, with DeepSeek-V4.1-Flash and GLM-5.3-Flash, so agreement with a reader other than propella-1-4b can be measured. The documents come from the corpora of the Dynaword project.

Loading

from datasets import load_dataset

train = load_dataset("danish-foundation-models/propella-bekko-data", "train", split="train")
test = load_dataset("danish-foundation-models/propella-bekko-data", "test-deepseek", split="train")
glm = load_dataset("danish-foundation-models/propella-bekko-data", "test-glm-5.3-flash", split="train")

row = test[0]
row["text"][:200]            # the document, as the annotators saw it
row["content_quality"]       # what propella-1-4b said
row["ref_content_quality"]   # what DeepSeek-V4.1-Flash said

train is 3.2 GB of text; add streaming=True if you only want to pass over it.

Motivation

Curating a pretraining corpus means reading every document: whether it is complete and readable, whether it is educational or safe, and what it is about. propella-1-4b does that reading, but at a fraction of a document per second on a single GPU, which does not scale to corpora of hundreds of millions of documents. These annotations are the distillation targets that turned that judgement into models that annotate hundreds of documents per second on the same hardware.

Composition

train holds the distillation targets: 844,912 documents with their propella-1-4b annotations. test-deepseek and test-glm-5.3-flash hold the same 1,000 held-out documents, annotated by the teacher and by one of the two reference annotators.

configuration	rows	labels	purpose
train	844,912	propella-1-4b	distillation targets
test-deepseek	1,000	propella-1-4b + DeepSeek-V4.1-Flash	agreement with an independent annotator
test-glm-5.3-flash	300	propella-1-4b + GLM-5.3-Flash	the same, against a second annotator

The held-out documents were sampled once, across 15 domains and 20 languages, before any training pool was built, and are excluded from train by source and document id and by text hash, so no test document appears in the training data. The GLM split is the 300-document subset of those 1,000 that the second annotator covered. The 1,000-document GLM run is not published because the endpoint returned empty answers for all but a handful of documents.

Documents come from five Dynaword corpora and from two web and PDF corpora annotated by the same project:

source	rows in train	this release
danish-dynaword	147,392	51 subsets
norwegian-dynaword	108,164	36 subsets
dutch-dynaword	83,337	20 subsets
icelandic-dynaword	56,020	16 subsets
faroese-dynaword	29,999	6 subsets
FineWeb-2	120,000	6 languages
FinePDFs	300,000	15 languages

The Dynaword rows carry their corpus' annotations and text, joined on subset and document id. The FineWeb-2 and FinePDFs rows take annotations from openeurollm/propella-annotations and text from the source corpus, joined on the document id. subset names the Dynaword subset (adl, lovdata, igc-law, and so on) or, for the two web and PDF corpora, their language partition (deu_Latn, swe_Latn). lang follows each corpus' own convention: bare ISO 639-3 for the Dynawords (dan, nor, isl, fao, nld) and xxx_Latn for FineWeb-2 and FinePDFs. Domain coverage follows from that: 300,000 PDF documents, 153,117 web, 143,095 legal, 60,638 encyclopedic, and the remaining 188,062 spread over news, books, conversation, social media and smaller domains.

Two things are worth knowing before using train. 53,761 rows (6.4%) carry no annotation, because the corpus subset they come from had none; they are kept as they are. Two (source, id) pairs appear twice.

Labels

The 18 properties of the propella-1 label space. Values are the label names themselves; multi-label properties hold a list of names, and ordinal properties are listed in level order.

category	property	type	values	meaning
Core content	content_integrity	ordinal, 4	complete → severely_degraded	completeness and technical quality
	content_ratio	ordinal, 5	complete_content → minimal_content	substantive content vs. navigation and boilerplate
	content_length	ordinal, 4	substantial → minimal	amount of substantive content
Classification	content_type	multi-label, 18	analytical, instructional, reference, procedural, legal_document, news_report, boilerplate, ...	functional purpose of the document
	business_sector	multi-label, 37	academic_research, education_sector, technology_software, financial_services, healthcare_medical, ...	industry relevance
	technical_content	multi-label, 7	code_heavy, math_heavy, scientific, data_heavy, engineering, basic_technical, non_technical	kind and intensity of specialised knowledge
Quality and value	content_quality	ordinal, 5	excellent → unacceptable	overall writing and presentation
	information_density	ordinal, 5	dense → empty	useful information relative to redundancy
	educational_value	ordinal, 5	high → none	potential for teaching and learning
	reasoning_indicators	ordinal, 5	analytical → none	presence of reasoning and analysis
Audience and purpose	audience_level	ordinal, 6	expert → children	target sophistication
	commercial_bias	ordinal, 5	none → pure_marketing	commercial influence on objectivity
	time_sensitivity	ordinal, 4	evergreen → time_sensitive	how quickly the value decays
Safety and compliance	content_safety	ordinal, 5	safe → illegal	inappropriate or harmful content
	pii_presence	binary	no_pii, contains_pii	personally identifiable information
Geography	regional_relevance	multi-label, 14	european, north_american, east_asian, global, culturally_neutral, ...	regional and cultural context
	country_relevance	multi-label, 749	denmark, norway, netherlands, germany, united_states, supranational, none, ...	country-level relevance
Classification	one_sentence_description	free text	—	neutral one-sentence summary

one_sentence_description is not predicted by the released annotators, which classify rather than generate. In the test configurations every property appears twice: unprefixed as propella-1-4b's label, and prefixed ref_ as the independent annotator's, with reference_annotator naming that annotator. Definitions and the full sets of values are in the propella-1 property reference.

Collection

Documents were sampled from whole parquet row groups, shuffled, at most 100 documents per row group, so that a sample is spread across a corpus rather than concentrated where the file begins. Caps were 5,000 documents per Dynaword subset and 20,000 per language for FineWeb-2 and FinePDFs.

train re-annotates nothing: its labels were produced by propella-1-4b for the Dynawords and for the two web and PDF corpora, and are read from where they are published. The reference annotations were made for this dataset. Each held-out document was sent to DeepSeek-V4.1-Flash or GLM-5.3-Flash with propella's annotator prompt and its JSON response schema, at temperature 0.0, and the answer was parsed without further editing. There are no human annotations anywhere in the dataset.

Preprocessing

text follows the model's view of a document: up to 50,000 characters from its start plus 10,000 from its end, joined by a \n[...]\n marker, which is what propella-1-4b annotated. n_chars and n_words give the full document length, and text_sha1 is the hash of the full, whitespace-normalised text, so every row can be traced back to its document in the source corpus and checked.

column	type	meaning
source, subset	string	corpus and subset (or language partition)
lang, domain	string	language and domain as labelled in the source corpus
id	string	document id within source and subset
text	string	document text in the model's 50k + 10k character view
text_sha1	string	sha1 of the whitespace-normalised full text
n_chars, n_words	int64	full-document length in characters and words
18 label columns	string or list of strings	propella-1-4b annotations
reference_annotator	string	null in train, the model id in the test configurations
18 ref_ columns	string or list of strings	the reference annotator's labels

Distribution

Nine parquet shards for train and one file per test configuration, 3.2 GB in total, zstd-compressed, public and ungated. The scripts that build the split files and the reference annotations are in scripts/, together with the commands to regenerate both.

Citation

This work was done as part of the Danish Foundation Models to curate data for the Dynawords. If you use the dataset, please cite the Dynaword paper (arXiv:2508.02271) and the project:

@inproceedings{dynaword,
    title = {Dynaword: {From} {One}-shot to {Continuously} {Developed} {Datasets}},
    booktitle = {Proceedings of the {Language} {Resources} and {Evaluation} {Conference} ({LREC} 2026)},
    author = {Enevoldsen, Kenneth and Jensen, Kristian Nørgaard and Kostkan, Jan and Szabó, Balázs and Kardos, Márton and Vad, Kirsten and Núñez, Andrea Blasi and Barmina, Gianluca and Nielsen, Jacob and Larsen, Rasmus and Vahlstrup, Peter and Dalum, Per Møldrup and Elliott, Desmond and Galke, Lukas and Schneider-Kamp, Peter and Nielbo, Kristoffer},
    year = {2026},
}
@misc{enevoldsenDanishFoundationModels2023,
    title = {Danish {Foundation} {Models}},
    url = {http://arxiv.org/abs/2311.07264},
    publisher = {arXiv},
    author = {Enevoldsen, Kenneth and Hansen, Lasse and Nielsen, Dan S. and Egebæk, Rasmus A. F. and Holm, Søren V. and Nielsen, Martin C. and Bernstorff, Martin and Larsen, Rasmus and Jørgensen, Peter B. and Højmark-Bertelsen, Malte and Vahlstrup, Peter B. and Møldrup-Dalum, Per and Nielbo, Kristoffer},
    year = {2023},
}

---

A Danish Foundation Models dataset