danish-foundation-models/icelandic-dyna-instruct

🧨 Icelandic dyna-instruct

<!-- START README TABLE --> | | | | ------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Version | 0.1.0 (Changelog) | | Language | Icelandic (isl) | | License | Openly Licensed, see individual datasets | | Models | For models trained on this data see danish-foundation-models | | Contact | If you have questions about this project please create an issue here | <!-- END README TABLE -->

Table of Contents

• —🧨 Icelandic dyna-instruct
• —Dataset Description
• —Dataset Summary
• —Loading the dataset
• —Languages
• —Task Types
• —Licensing
• —Dataset Structure
• —Data Instances
• —Data Fields
• —Data Splits
• —Dataset Creation
• —Curation Rationale
• —Source Data
• —Data Collection and Processing
• —Dataset Statistics
• —Contributing to the dataset
• —Citation Information
• —License Information

Dataset Description

<!-- START-DESC-STATS -->

• —Number of samples: 8.11K
• —Number of tokens (Llama 3): 7.09M
• —Average conversation length in tokens (min, max): 874.89 (182, 1.39K)
• —Average number of turns (min, max): 2.0 (2, 2) <!-- END-DESC-STATS -->

Dataset Summary

Icelandic dyna-instruct is a collection of instruction-tuning datasets from various sources. All datasets are openly licensed and deemed permissible for training large language models. The dataset is used to develop the DFM models.

Icelandic dyna-instruct is continually developed using the Dynaword approach, meaning it 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/icelandic-dyna-instruct"
ds = load_dataset(name, split="train")
sample = ds[0]

Or stream it:

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

Load a single subset:

ds = load_dataset(name, "dynaword-reverse-instruct", split="train")

Pin a revision for reproducibility:

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

Languages

This dataset includes multiple languages. The following figures give an overview of the languages in the dataset.

<img src="./images/task_distribution.png" width="800"/>

Task Types

<!-- START-TASK TABLE --> | Task | Sources | N. Tokens | |:----------|:----------------------------|:------------| | knowledge | [dynaword-reverse-instruct] | 7.09M | | Total | | 7.09M |

[dynaword-reverse-instruct]: data/dynaword-reverse-instruct/datasheet.md <!-- END-TASK TABLE -->

The following task types are used:

• —general — diverse SFT mixtures covering many capabilities
• —translation — Icelandic↔English translation
• —reasoning — logical and mathematical reasoning with chain-of-thought
• —refusal — safety refusals for harmful requests
• —code — code generation and agentic code tasks
• —tool_use — function calling and tool selection
• —instruction_following — instruction-following benchmarks
• —knowledge — knowledge-grounded generation
• —qa — question answering

Licensing

<!-- START-LICENSE TABLE --> | License | Sources | N. Tokens | |:----------|:----------------------------|:------------| | Unknown | [dynaword-reverse-instruct] | 7.09M | | Total | | 7.09M |

[dynaword-reverse-instruct]: data/dynaword-reverse-instruct/datasheet.md <!-- END-LICENSE TABLE -->

Dataset Structure

Data Instances

<!-- START-SAMPLE -->

{
  "id": "dynaword-reverse-instruct_00000",
  "messages": [
    {
      "role": "user",
      "content": "Hvað er ljós, hvernig er eðli þess lýst í eðlisfræði og hvað einkennir sýnilegt ljós?"
    },
    {
      "role": "assistant",
      "content": "Ljós Ljós er rafsegulbylgjur innan ákveðins tíðnisviðs, en oftast er átt við það tíðnisvið sem mannsaugað greinir. Við n[...]"
    }
  ],
  "source": "dynaword-reverse-instruct",
  "added": "2026-09-01",
  "created": "2026-01-01, 2026-12-31",
  "language": [
    "isl"
  ],
  "task": "knowledge",
  "token_count": 588
}

Data Fields

An entry in the dataset consists of the following fields:

• —id (str): A unique identifier for each sample.
• —messages (list[dict]): The conversation turns — see the message format section in the main README.
• —source (str): The source dataset (see Source Data).
• —added (str): The date when the sample was added to this collection.
• —created (str): The approximate date range when the source data was created, as "YYYY-MM-DD, YYYY-MM-DD" (start, end). Year-level granularity is used when the exact date is unknown.
• —language (list[str]): The language(s) of the sample as ISO 639-3 codes (e.g. ["isl"], ["eng"], ["isl", "eng"]).
• —task (str): The task type of the sample.
• —token_count (int): The total number of tokens across all messages, computed using the Llama 3 tokenizer. <!-- END-SAMPLE -->

Data Splits

The entire corpus is provided in the train split.

Dataset Creation

Curation Rationale

These datasets were collected and curated to make openly licensed instruction-tuning data available for post-training large language models. Icelandic dyna-instruct forms the post-training complement to Icelandic Dynaword, a continuously developed collection of Icelandic pre-training text. Both are part of the Danish Foundation Models initiative. While collected with model training in mind, the data may also be useful for evaluation and research on Icelandic language instruction-following.

Source Data

Below is an overview of the sources with their individual licenses.

<details> <summary><b>Overview Table (click to unfold)</b></summary>

<!-- START-MAIN TABLE --> | Source | Description | Task | N. Tokens | License | |:----------------------------|:----------------------------------------------------------------------|:----------|:------------|:----------| | [dynaword-reverse-instruct] | Icelandic instruction-following pairs derived from Icelandic Dynaword | knowledge | 7.09M | [Unknown] | | Total | | | 7.09M | |

[dynaword-reverse-instruct]: data/dynaword-reverse-instruct/datasheet.md

[CC-0]: https://creativecommons.org/publicdomain/zero/1.0/legalcode.en [CC-BY-SA 4.0]: https://creativecommons.org/licenses/by-sa/4.0/deed.en [CC-BY 4.0]: https://creativecommons.org/licenses/by/4.0/deed.en [Apache 2.0]: https://www.apache.org/licenses/LICENSE-2.0 [Unknown]: ./data/dynaword-reverse-instruct/datasheet.md#license-information <!-- END-MAIN TABLE -->

</details>

Data Collection and Processing

Icelandic dyna-instruct is continually developed. Data collection and processing details are documented in the individual datasheets linked in the table above, and where possible in a reproducible data/{dataset}/create.py script.

In addition to dataset-specific processing, automated quality checks ensure correct column formatting, unique IDs, and no duplicate conversations.

Dataset Statistics

The following plot shows the approximate creation date range of each source:

<p align="center"> <img src="./images/creation_dates.png" width="800" /> </p>

The following plots give an overview of conversation length and turn distributions across sources:

<div style="display: flex; gap: 20px; align-items: flex-start;"> <p align="center"> <img src="./images/datasetsizeplot.svg" width="400" /> </p> <p align="center"> <img src="./images/datasetturnsplot.svg" width="400" /> </p> </div>

Contributing to the dataset

Icelandic dyna-instruct is continually developed.

We welcome contributions including new sources, improved filtering, and other enhancements. See CONTRIBUTING.md to get started.

Citation Information

<!-- TODO: replace with actual release citation once published --> If you use this work, please cite:

@article{enevoldsen2023danish,
  title={Danish Foundation Models},
  author={Enevoldsen, Kenneth and Hansen, Lasse and Nielsen, Dan S. and
          Egeb{\ae}k, Rasmus A. F. and Holm, S{\o}ren V. and Nielsen, Martin C. and
          Bernstorff, Martin and Larsen, Rasmus and J{\o}rgensen, Peter B. and
          H{\o}jmark-Bertelsen, Malte and Vahlstrup, Peter B. and
          M{\o}ldrup-Dalum, Per and Nielbo, Kristoffer},
  year={2023},
  journal={arXiv preprint arXiv:2311.07264},
  url={https://arxiv.org/abs/2311.07264}
}

License Information

The license for each constituent dataset is listed in the Source Data table. The collection itself (metadata, quality control, etc.) is licensed under CC-BY 4.0.

Notice and takedown policy

We redistribute files under licenses permitting such redistribution. If you have concerns about licensing or copyright, please contact us.

<h3 style="display: flex; align-items: center;"> <a href="https://www.foundationmodels.dk"> <img src="./docs/icon.png" width="30" style="margin-right: 10px;" /> </a> A&nbsp;<a href="https://www.foundationmodels.dk">Danish Foundation Models</a>&nbsp;dataset </h3>