CoolFace
Datasetpublic

pdelobelle/staatsblad-synth-nl

Synthetic Dutch from the Belgisch Staatsblad Diverse, fluent Dutch pretraining text synthesized from guust-franssens/belgisch-staatsblad (CC0, Belgian official-gazette filings). Adds Belgium/Flanders coverage to Dutch LM pretraining mixes, where clean Belgian-Dutch prose is otherwise scarce. The source text is noisy OCR from scanned PDFs, but its metadata (company, juridical form, act type, city, date) is clean. A local LLM (google/gemma-2-9b-it) "launders" the OCR + metadata… See the full description on the dataset page: https://huggingface.co/datasets/pdelobelle/staatsblad-synth-nl.

sourceHugging Facecc0-1.0updated 2mo agoView on Hugging Face
0likes416downloads
Dataset Card

Synthetic Dutch from the Belgisch Staatsblad

Diverse, fluent Dutch pretraining text synthesized from `guust-franssens/belgisch-staatsblad` (CC0, Belgian official-gazette filings). Adds Belgium/Flanders coverage to Dutch LM pretraining mixes, where clean Belgian-Dutch prose is otherwise scarce.

The source text is noisy OCR from scanned PDFs, but its metadata (company, juridical form, act type, city, date) is clean. A local LLM (`google/gemma-2-9b-it`) "launders" the OCR + metadata into fluent Dutch, with diversity enforced Cosmopedia/SYNTH-style: each generation samples an independent combination of conditional features (content type, audience, format, tone, length, focus). French/German filings are rendered into Dutch.

Statistics

texts185,640
tokens (pdelobelle/robbert-v2-dutch-base)~49.3M (~265/text)
source coverage123,741 / 784,358 filings (15.8%)
content mixfacts 40% · paraphrase 25% · qa 20% · summary 15%
source language~66% NL · ~34% FR→NL

Schema

text (the synthetic Dutch) plus conditioning/provenance fields: content_type, audience, format, tone, length, focus, source_pubid, source_company, source_form, source_act, source_city, source_lang, model.

Intended use & limitations

Intended as diverse pretraining filler, not a knowledge base. It is model output, not verified fact:

  • —Content is dominated by routine company filings (incorporations, statute changes, capital movements) — broad but repetitive subject matter.
  • —Geography and numbers occasionally err despite anti-fabrication prompting.
  • —Do not treat any specific claim as authoritative.

Generation & licensing note

The source gazette data is CC0. This synthetic text was generated with google/gemma-2-9b-it; using a general-purpose model to synthesize data is a gray area for strict "provenance-clean" pretraining goals. Review the license field and the generator disclosure before relying on this for a compliance-constrained pipeline.