danish-foundation-models/norwegian-dyna-instruct

🧨 Norwegian dyna-instruct

Version	0.1.0 (changelog)
Languages	Norwegian Bokmål (nob), Norwegian Nynorsk (nno), and English (eng) translation input
License	Mixed open licenses; see the table below
Sources	Five datasets (source cards)

Dataset Description

• Number of samples: 14.40K
• Number of tokens (Llama 3): 6.27M
• Average conversation length in tokens (min, max): 435.63 (4, 8.92K)
• Average number of turns (min, max): 2.13 (2, 3)

Dataset Summary

Norwegian dyna-instruct is a collection of openly licensed Norwegian instruction-tuning datasets. The five subsets share a common chat format and can be loaded together or separately. Each subset has its own card describing its source, preparation, and license.

Sources are included when their origin and creation are documented well enough to understand the resulting training examples. The source cards document each included source, its provenance, and the checks applied during preparation.

Benchmark note: fleurs-alpaca-en-no contains FLEURS training text drawn from FLORES-101 dev and devtest. Exclude this subset when evaluating on FLEURS or FLORES.

Loading the dataset

Load the full collection with:

from datasets import load_dataset

name = "danish-foundation-models/norwegian-dyna-instruct"
ds = load_dataset(name, split="train")
sample = ds[0]

Load a single subset by name:

ds = load_dataset(name, "nb-samtale-pairs", split="train")

Pass revision= to load_dataset to pin a particular release.

Languages

Rows use ISO 639-3 codes: nob for Bokmål, nno for Nynorsk, and eng for the English side of translation examples.

Task Types

Task	Sources	N. Tokens
general	nb-samtale-pairs, magpie-qwen3-bokmaal	1.93M
instruction_following	magpie-qwen3-bokmaal, reasoning-norwegian	1.74M
qa	norquad-wikipedia	1.44M
reasoning	magpie-qwen3-bokmaal	853.65K
code	magpie-qwen3-bokmaal	189.17K
translation	fleurs-alpaca-en-no	115.89K
Total		6.27M

Licensing

License	Sources	N. Tokens
Apache 2.0	magpie-qwen3-bokmaal	2.34M
CC BY-SA 4.0	fleurs-alpaca-en-no, reasoning-norwegian	1.74M
CC BY-SA 3.0	norquad-wikipedia	1.44M
CC0	nb-samtale-pairs	740.05K
Total		6.27M

The subsets keep their original licenses. Check the table and the linked subset card before using or redistributing the data. Attribution files for ShareAlike subsets are stored next to the data and are not returned by load_dataset.

Dataset Structure

Data Instances

{
  "id": "nb-samtale-pairs_nb-1_0000",
  "messages": [
    {
      "role": "user",
      "content": "Hallo og velkommen hit til Nasjonalbiblioteket. Det er kulturnatt, og det skjer en bråte ting over hele byen, og det gjø[...]"
    },
    {
      "role": "assistant",
      "content": "Tusen takk Lars. Vi er altså mer spennende enn kokte poteter. Det var godt å høre. Det var noen som hviska bak her, er v[...]"
    }
  ],
  "source": "nb-samtale-pairs",
  "added": "2026-09-03",
  "created": "2017-01-01, 2021-12-31",
  "language": [
    "nob"
  ],
  "task": "general",
  "token_count": 1227
}

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, each with a role and a text content field.
• 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. ["nob"], ["nno"], ["nob", "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.

Data Splits

The collection has one train split. Each subset is also available as a named configuration.

Dataset Creation

Curation Rationale

The collection puts reusable Norwegian post-training data into one consistent format. It complements norwegian-dynaword and follows the continuously developed Dynaword approach.

Source Data

Included-source overview (click to unfold)

Source	Description	Task	N. Tokens	License
magpie-qwen3-bokmaal	Synthetic Bokmål conversations generated by Qwen3-235B-A22B with Magpie	general, reasoning, code, instruction_following	2.34M	Apache 2.0
reasoning-norwegian	Punctuation and capitalization corrections for Norwegian Wikipedia passages	instruction_following	1.63M	CC BY-SA 4.0
norquad-wikipedia	Norwegian Wikipedia questions and answers in chat format	qa	1.44M	CC BY-SA 3.0
nb-samtale-pairs	Consecutive turns from the NB Samtale transcripts	general	740.05K	CC0
fleurs-alpaca-en-no	English-to-Bokmål translation pairs from FLEURS	translation	115.89K	CC BY-SA 4.0
Total			6.27M

The linked source cards record upstream revisions, creation methods, licensing evidence, attribution, and preparation steps for each included subset.

Data Collection and Processing

Each subset is converted to the shared chat format by a create.py script with pinned source revisions. Shared checks validate the schema, remove empty or duplicate conversations, and calculate token counts. The subset cards describe their individual processing steps.

Dataset Statistics

The following plots are regenerated from the checked-in Parquet files:

Contributing to the dataset

See CONTRIBUTING.md to propose a new source or improve an existing one.

Citation Information

Please cite the original sources listed in the subset cards. For the collection methodology, cite Dynaword:

@article{enevoldsen2025dynaword,
  title = {Dynaword: From One-shot to Continuously Developed Datasets},
  author = {Enevoldsen, Kenneth and others},
  year = {2025},
  journal = {arXiv preprint arXiv:2508.02271},
  url = {https://arxiv.org/abs/2508.02271}
}

License Information

There is no single content license for the collection. Each subset keeps the license shown in the table above. Code and documentation created for this repository are licensed under Apache 2.0.

Notice and takedown policy

Open an issue if you find a rights or privacy concern, and identify the affected source or rows where possible.