FO-Instruct: Faroese Instruction Dataset

FO-Instruct is a Faroese instruction dataset containing 571 instruction–response pairs. The dataset was created to support the development, training, and evaluation of language models that understand and generate Faroese.

Each example consists of an instruction or question in Faroese and a corresponding response in Faroese.

The dataset was developed at the Centre for Language Technology (Máltøknidepilin), University of the Faroe Islands (Fróðskaparsetur Føroya) by Iben Nyholm Debess, Uni Johannesen, and Birita Dal-Christiansen.

Dataset structure

The dataset contains 571 examples with the following fields:

Field	Description
contributor_id	Pseudonymous identifier for the contributor who created the example
instruction	An instruction or question written in Faroese
output	The corresponding response written in Faroese

The contributor_id field preserves information about the origin of individual examples. The three contributors are represented as contributor_1, contributor_2, and contributor_3.

The contributors created their examples independently, and the full collections were not systematically cross-reviewed for duplicate or closely related topics. Retaining a contributor identifier therefore makes it possible for users to identify examples originating from the same contributor and to take possible topical overlap or contributor-specific patterns into account when analysing or using the dataset.

The contributor identifiers are pseudonymous within the row-level data and are not intended to map individual examples to named contributors.

Example

{
  "contributor_id": "contributor_1",
  "instruction": "Hvat eitur hægsta fjall í Føroyum, og hvussu høgt er tað?",
  "output": "Hægsta fjallið eitur Slættaratindur og er 880 metrar høgt."
}

Another example:

{
  "contributor_id": "contributor_1",
  "instruction": "Hvat er slag av gróti er mest av í Føroyum?",
  "output": "Í Føroyum er mest basalt, ið stavar frá eldgosum."
}

Content

The dataset contains a diverse collection of Faroese instruction–response pairs covering both Faroese-specific knowledge and more general topics.

A substantial part of the dataset focuses on the Faroe Islands, including:

• Faroese history and society
• geography, geology and nature
• transport and infrastructure
• politics and public institutions
• media and religion
• industry, fisheries and the economy
• Faroese culture, traditions and national symbols
• literature, authors and other notable Faroese people
• traditional Faroese clothing
• sports and Faroese athletes

The dataset also contains extensive material about the Faroese language, including:

• grammar and word classes
• spelling and orthography
• vocabulary and etymology
• loanwords and inherited words
• language history
• Faroese dialects and regional variation
• the development of the Faroese written language

Other examples cover the Nordic countries and neighbouring regions, including Denmark, Iceland, Norway, Sweden, Finland, Greenland, Shetland and Orkney, as well as general European geography, history and culture.

There are also instruction–response pairs concerning areas such as literature and children's literature, law, taxation, mathematics, science, health, food, travel, sports, everyday advice and general knowledge.

The dataset therefore combines factual question answering with more open-ended instructions, explanations and recommendations, providing a varied resource for developing and evaluating Faroese instruction-following language models.

Intended uses

FO-Instruct can be used for tasks such as:

• instruction tuning
• supervised fine-tuning
• question answering
• evaluation of Faroese instruction-following capabilities
• Faroese text generation
• research on language technology for low-resource languages

The dataset may also be combined with other Faroese or multilingual datasets when training or evaluating language models.

The contributor_id field can additionally be used when analysing contributor-specific patterns or investigating whether closely related examples originate from the same or different contributors.

Language

The dataset is entirely in Faroese.

The ISO 639-1 language code for Faroese is fo.

Data format

The dataset is distributed as JSON Lines (JSONL), with one example per line.

{"contributor_id":"contributor_1","instruction":"Hvat er slag av gróti er mest av í Føroyum?","output":"Í Føroyum er mest basalt, ið stavar frá eldgosum."}
{"contributor_id":"contributor_1","instruction":"Nær vórðu Føroyar til jarðfrøðiliga?","output":"Jarðfrøðiliga vórðu Føroyar til í tertiertíðini, fyri umleið 55-60 milliónum árum síðani."}

The dataset file is located at:

data/train.jsonl

Loading the dataset

The dataset can be loaded using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("maltoknidepilin/fo-instruct")

print(dataset["train"][0])

An individual example will have the following structure:

{
    "contributor_id": "contributor_1",
    "instruction": "...",
    "output": "..."
}

Dataset creation

FO-Instruct was developed at the Centre for Language Technology (Máltøknidepilin) at the University of the Faroe Islands (Fróðskaparsetur Føroya).

The instruction–response pairs were manually created by three contributors.

Contributors worked independently on their respective sets of examples. The complete contributions were not systematically cross-reviewed for duplicate or closely related topics. As a result, similar subjects may occur in examples created independently by different contributors.

To preserve this provenance without directly identifying the creator of each individual row, each example contains a pseudonymous contributor_id.

Creators

• Iben Nyholm Debess
• Uni Johannesen
• Birita Dal-Christiansen

Limitations

FO-Instruct is a relatively small dataset consisting of 571 examples. It is therefore not intended to provide comprehensive coverage of Faroese language, knowledge, subject areas, or instruction types.

Because the contributors created examples independently and the complete dataset was not systematically cross-reviewed for topical overlap, the dataset may contain instructions that address identical or closely related subjects.

There may also be contributor-specific differences in question formulation, topic selection, level of detail, and response style. The contributor_id field is retained so that users can take such patterns into account where relevant.

The responses in the dataset should not necessarily be regarded as authoritative sources of factual information. Users should evaluate the suitability and accuracy of the data for their particular application.

Some responses concern facts that may change over time, and users should therefore independently verify time-sensitive information when necessary.

License

FO-Instruct is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Citation

If you use FO-Instruct in research or other published work, please cite the dataset:

@dataset{fo_instruct_2026,
  title     = {FO-Instruct: Faroese Instruction Dataset},
  author    = {Debess, Iben Nyholm and Johannesen, Uni and Dal-Christiansen, Birita},
  publisher = {Centre for Language Technology, University of the Faroe Islands},
  year      = {2026}
}

About the Centre for Language Technology

The Centre for Language Technology (Máltøknidepilin) at the University of the Faroe Islands develops language resources and technologies for Faroese.

More information and Faroese language technology resources are available at:

https://mtd.setur.fo/