tvosch/GPT-NL-propella-annotations

GPT-NL Propella Annotations

Document-level quality and content annotations for the Dutch subset of GPT-NL/GPT-NL_Public_Corpus.

What is Propella?

Propella (ellamind/propella-1-4b) is a 4B-parameter language model fine-tuned from Qwen3 for document-level annotation of LLM pretraining data. Given a document in any language, it produces a structured JSON object with 18 quality, classification, and safety properties (see below for the features). These annotations can be used downstream to filter or weight training data, for example, removing low-quality documents, deduplicating by content type, or balancing domain coverage. See also openeurollm/propella-annotations for more info.

Dataset description

Each row corresponds to one Dutch document from the GPT-NL Public Corpus and contains propella's structured annotation alongside the original text. The annotation covers content quality, type, audience, safety, and several other dimensions useful for data curation and filtering.

Size: 31,291,097 documents Tokens: 50,297,442,743 tokens

Subsets

Subsets have been filtered base on the GPT-NL's original Dutch language filter as stated in the column language. That results in:

Subset	Config name	Documents
American-stories	american_stories	15
Auditdienst Rijk	auditdienst_rijk	555
Belgian Journal	belgian_journal	208,242
C5 Filtered	c5_filtered	63,519
CC-English-PD	cc_english_pd	1,329
CC-Eurovoc	cc_eurovoc	34,948
CC-German-PD	cc_german_pd	8,177
CC-Github Code	cc_github_code	38,083
CC-Loc-PD-Books	cc_loc_pd_books	14
DANS-KNAW	dans_knaw	52,880
De Rechtspraak	de_rechtspraak	918,634
Dienst Publiek en Communicatie	dienst_publiek_en_communicatie	127,715
Eurlex	eurlex	36,810
European Parliament	european_parliament	654
Koninklijke Bibliotheek	koninklijke_bibliotheek	1,571,895
Nationaal Archief	nationaal_archief	1,924,127
Naturalis	naturalis	2,652
Noord-Hollands Archief	noord_hollands_archief	38,737
Officiele Bekendmakingen	officiele_bekendmakingen	1,822,093
Openraadsinformatie	openraadsinformatie	2,712,533
PBL	pbl	341
Tweede Kamer	tweede_kamer	229,600
Utrechts Archief	utrechts_archief	525,886
Wikidata-Synth	wikidata_synth	14,582,582
Wikiwijs	wikiwijs	119,187
Woogle	woogle	4,088,931
YouTube-Commons-Synth	youtube_commons_synth	2,147,061
Zeeuws Archief	zeeuws_archief	33,897
Total		31,291,097

Note on source_id uniqueness

The source_id column is the original id field from GPT-NL_Public_Corpus. It is not a document-level unique identifier for the majority of the corpus.

source_id can be used together with dataset_name (available in the original corpus) to identify the sub-corpus a document comes from, but it does not uniquely identify individual documents for most sub-corpora. Use doc_id as the stable unique identifier for this dataset.

Annotation errors

A small fraction of documents (~239,000 out of 31M, <1%) could not be annotated due to model output validation failures. These rows have null values for all annotation columns and a non-null annotation_error string. Filter them with:

ds = ds.filter(lambda x: x["annotation_error"] is None)

Annotation setup

The model was served with vLLM (v0.19.0) inside an Apptainer container, with structured JSON output constrained by xgrammar:

vllm serve ellamind/propella-1-4b \
    --tensor-parallel-size 1 \
    --max-model-len 32768 \
    --max-num-seqs 256 \
    --quantization fp8 \
    --kv-cache-dtype fp8 \
    --async-scheduling \
    --enable-prefix-caching

Source corpus & license

Source: GPT-NL/GPT-NL_Public_Corpus
License: CC-BY-4.0

Columns

Column	Type	Description
doc_id	string	Deterministic UUID5 unique identifier, stable across re-runs. Derived from the source shard index and row position.
source_id	string	Original id value from GPT-NL_Public_Corpus. See note on uniqueness below.
dataset_name	string	Sub-corpus name from GPT-NL_Public_Corpus (e.g. C5 Filtered, Koninklijke Bibliotheek).
text	string	Document text (CC-BY-4.0, from GPT-NL_Public_Corpus).
content_integrity	string	Whether the text is complete and coherent.
content_ratio	string	Ratio of meaningful content vs. boilerplate/noise.
content_length	string	Qualitative length category.
one_sentence_description	string	One-sentence summary of the document.
content_type	list[string]	Content type tags (e.g. news, legal_document, boilerplate).
business_sector	list[string]	Relevant industry/sector tags.
technical_content	list[string]	Degree and type of technical content.
information_density	string	How information-dense the text is.
content_quality	string	Overall quality rating.
audience_level	string	Intended audience (e.g. general, expert).
commercial_bias	string	Degree of commercial intent.
time_sensitivity	string	Whether content is time-sensitive or evergreen.
content_safety	string	Safety classification.
educational_value	string	Educational value rating.
reasoning_indicators	string	Presence of reasoning or argumentation.
pii_presence	string	Whether the document contains personally identifiable information.
regional_relevance	list[string]	Geographic relevance tags.
country_relevance	list[string]	Specific country relevance tags.
annotation_error	string	Populated only when the annotation model failed for this document; all annotation columns will be null in that case.