danish-foundation-models/dutch-dynaword

🧨 Dutch Dynaword

Version	1.0.0 (Changelog)
Language	nld, Nederlands, Dutch
License	Openly Licensed, See the respective dataset
Models	For model trained used this data see danish-foundation-models
Contact	If you have question about this project please create an issue here

Dataset Description

• Number of samples: 560
• Number of tokens (Llama 3): 7.16M
• Average document length in tokens (min, max): 12.79K (448, 107.38K)

Dataset Summary

The Dutch dynaword is a collection of Dutch free-form text datasets from various domains. All of the datasets in Dutch Dynaword are openly licensed and deemed permissible for training large language models.

Dutch Dynaword is continually developed, which means that the dataset will actively be updated as new datasets become available. If you would like to contribute a dataset see the contribute section.

Loading the dataset

from datasets import load_dataset

name = "danish-foundation-models/dutch-dynaword"
ds = load_dataset(name, split = "train")
sample = ds[1] # see "Data Instances" below

or load it by streaming the data

ds = load_dataset(name, split = "train", streaming=True)
dataset_iter = iter(ds)
sample = next(iter(dataset_iter))

You can also load a single subset at a time:

ds = load_dataset(name, "auditdienstrijk", split = "train")

To allow filtering we additionally provide extensive annotations available through the meta config:

meta = load_dataset(name, "meta", split = "train")

For more on how to use the annotations see the annotations section.

As Dutch Dynaword is continually expanding and curated you can make sure that you get the same dataset every time by specifying the revision: You can also load a single subset at a time:

ds = load_dataset(name, revision="{desired revision}")

Languages

This dataset includes the following languages:

• Dutch (nld-Latn)

In addition it likely contains small amounts of English due to code-switching.

Language is denoted using BCP-47, using the langauge code ISO 639-3 and the script code ISO 15924. The third element denote the region variant.

Domains

This dynaword consist of data from various domains (e.g., legal, books, social media). The following table and figure give an overview of the relative distributions of these domains. To see a full overview of the source check out the source data section

Domain	Sources	N. Tokens
Legal	auditdienstrijk	7.16M
Total		7.16M

Annotation Overview

Each document in Dutch Dynaword comes with annotations describing its content, such as content quality, information density, and educational value. Each bar shows the share of documents at each level of one annotation, from worst (light) to best (dark). The same plot is available for every source in its datasheet, and the counts behind it are stored in descriptive_stats.json under annotations. To learn more, see the annotations section.

Licensing

The following gives an overview of the licensing in the Dynaword. To get the exact license of the individual datasets check out the overview table. These license is applied to the constituent data, i.e., the text. The collection of datasets (metadata, quality control, etc.) is licensed under CC-0.

License	Sources	N. Tokens
CC-0	auditdienstrijk	7.16M
Total		7.16M

Dataset Structure

The dataset contains text from different sources which are thoroughly defined in Source Data.

Data Instances

Each entry in the dataset consists of a single text with associated metadata

{
  "id": "auditdienstrijk_0",
  "text": "Auditrapport\n2021\nMilou Meyer (I) Auditrapport 2021\nMilou Meyer (I)\n15 maart 2022\nKenmerk\n42490419\nI[...]",
  "source": "auditdienstrijk",
  "added": "2026-09-01",
  "created": "2016-01-01, 2025-12-31",
  "token_count": 7818
}

Data Fields

An entry in the dataset consists of the following fields:

• id (str): A unique identifier for each document.
• text (str): The content of the document.
• source (str): The source of the document (see Source Data).
• added (str): The date when the document was added to this collection.
• created (str): The date range when the document was originally created.
• token_count (int): The number of tokens in the sample computed using the Llama 3 tokenizer.

Data Splits

The entire corpus is provided in the train split.

Dataset Creation

Curation Rationale

These datasets were collected and curated with the intention of making openly licensed Dutch data available. While this was collected with the intention of developing language models it is likely to have multiple other uses such as examining language development and differences across domains.

Annotations

Synthetic metadata is stored as data/{dataset}/metadata.parquet. These annotations were generated with ellamind/propella-1-4b and include fields for content type, quality, safety, audience level, educational level, PII presence, regional relevance and more.

The metadata rows include dataset and id, so a subset can be filtered and joined with the corpus rows:

from datasets import load_dataset

name = "danish-foundation-models/dutch-dynaword"
texts = load_dataset(name, "auditdienstrijk", split="train")
meta = load_dataset(name, "meta", split="train").filter(
    lambda row: row["dataset"] == "auditdienstrijk"
)

texts_df = texts.to_pandas()
meta_df = meta.to_pandas()
adr_with_meta = texts_df.merge(meta_df, on="id", how="left")

Source Data

Below follows a brief overview of the sources in the corpus along with their individual license. To get more information about the individual dataset click the hyperlink in the table.

Overview Table (click to unfold)

You can learn more about each dataset by pressing the link in the first column.

Source	Description	Domain	N. Tokens	License
auditdienstrijk	Audit reports from the Auditdienst Rijk	Legal	7.16M	CC-0
Total			7.16M

Data Collection and Processing

Dutch Dynaword is continually developed, which means that the dataset will actively be updated as new datasets become available. This means that the size of Dynaword increases over time as seen in the following plot:

The data collection and processing varies depending on the dataset and is documentationed the individual datasheets, which is linked in the above table. If possible the collection is documented both in the datasheet and in the reproducible script (data/{dataset}/create.py).

In addition to data specific processing we also run a series automated quality checks to ensure formatting (e.g. ensuring correctly formatted columns and unique IDs), quality checks (e.g. duplicate and empty string detection) and datasheet documentation checks. These checks are there to ensure a high quality of documentation and a minimal level of quality. To allow for the development of novel cleaning methodologies we do not provide more extensive cleaning.

Dataset Statistics

The following plot(s) are intended to give an overview of docuements length in the various sources.

Contributing to the dataset

We welcome contributions to the dataset, including new sources, improved data filtering, and other enhancements. To get started on contributing, please see the contribution guidelines

Citation Information

If you use this work, please cite the scientific article, we recommend citing the following:

Enevoldsen, K.C., Jensen, K.N., Kostkan, J., Szab'o, B.I., Kardos, M., Vad, K., Heinsen, J., N'unez, A.B., Barmina, G., Nielsen, J., Larsen, R., Vahlstrup, P.B., Dalum, P.M., Elliott, D., Galke, L., Schneider-Kamp, P., & Nielbo, K.L. (2025). Dynaword: From One-shot to Continuously Developed Datasets.

@article{enevoldsen2025dynaword,
  title={Dynaword: From One-shot to Continuously Developed Datasets},
  author={Enevoldsen, Kenneth and Jensen, Kristian N{\o}rgaard and Kostkan, Jan and Szab{\'o}, Bal{\'a}zs and Kardos, M{\'a}rton and Vad, Kirten and N{\'u}{\~n}ez, Andrea Blasi and Barmina, Gianluca and Nielsen, Jacob and Larsen, Rasmus and others},
  journal={arXiv preprint arXiv:2508.02271},
  year={2025}
}

Additionally, we recommend citing the relevant source datasets as well. See the individual datasheets for more information.

License information

The license for each constituent dataset is supplied in the Source data table. This license is applied to the constituent data, i.e., the text. The collection of datasets (metadata, quality control, etc.) is licensed under CC-0.

Personal and Sensitive Information

As far as we are aware the dataset does not contain information identifying sexual orientation, political beliefs, religion, or health connected with utterer ID. In case that such information is present in the data we have been removed utterer information from social media content.

Bias, Risks, and Limitations

Certain works in this collection are historical works and thus reflect the linguistic, cultural, and ideological norms of their time. As such, it includes perspectives, assumptions, and biases characteristic of the period.

Notice and takedown policy

We redistribute files shared with us under a license permitting such redistribution. If you have concerns about the licensing of these files, please contact us. If you consider that the data contains material that infringe your copyright, please:

• Clearly identify yourself with detailed contact information such as an address, a telephone number, or an email address at which you can be contacted.
• Clearly reference the original work claimed to be infringed
• Clearly identify the material claimed to be infringing and information reasonably sufficient to allow us to locate the material. You can contact us through this channel. We will comply with legitimate requests by removing the affected sources from the next release of the corpus

---

A Danish Foundation Models dataset