birgermoell/swedish-runestone-inscriptions

Scandinavian Runestone Inscriptions

A multimodal dataset of 2,615 ancient Scandinavian runic inscriptions paired with photographs of the runestones. Each entry contains a photograph of a runestone along with its scholarly transliteration (in runic), Old Norse normalization, and English translation.

This dataset is designed for evaluating Vision-Language Models (VLMs) on the task of reading and transcribing runic inscriptions from stone monument photographs -- a challenging task that combines historical epigraphy with computer vision.

Dataset Description

Overview

The dataset contains 2,615 entries from Scandinavian runic inscriptions (runestones, rock carvings, church walls, and other monuments), spanning from the Migration Period (ca. 400 CE) through the Medieval period (ca. 1500 CE), with the majority dating to the Viking Age (ca. 800--1100 CE). Each entry pairs a photograph from Wikimedia Commons with ground-truth scholarly data from the Scandinavian Runic Text Database (Rundata), maintained by Uppsala University.

The inscriptions are primarily written in Younger Futhark runes (the runic alphabet used in Scandinavia during the Viking Age), with some in Elder Futhark (Migration/Vendel Period) and medieval runes. Most Viking Age inscriptions are memorial texts commemorating deceased relatives ("X raised this stone in memory of Y, his/her father/mother/brother").

Key Statistics

Statistic	Value
Total entries	2,615
Evaluation entries	2,610
Few-shot examples	5
Countries covered	3 (Sweden, Norway, Denmark)
Provinces covered	17
Period range	ca. 400--1600 CE
Predominant period	Viking Age (~56%)
Min inscription length	10 rune characters
Max inscription length	801 rune characters
Mean inscription length	52.8 characters
Median inscription length	46 characters
Entries with normalization	2,615 (100%)
Entries with English translation	2,615 (100%)

Inscription Length Distribution

Category	Count	Percentage
Short (10--19 chars)	476	18.2%
Medium (20--60 chars)	1,330	50.9%
Long (> 60 chars)	809	30.9%

Province Distribution

Province	Count	%	Country
Uppland	786	30.1%	Sweden
Norway	420	16.1%	Norway
Denmark	287	11.0%	Denmark
Södermanland	282	10.8%	Sweden
Gotland	252	9.6%	Sweden
Västergötland	147	5.6%	Sweden
Östergötland	123	4.7%	Sweden
Småland	117	4.5%	Sweden
Unknown	80	3.1%	--
Öland	32	1.2%	Sweden
Västmanland	23	0.9%	Sweden
Närke	20	0.8%	Sweden
Gästrikland	14	0.5%	Sweden
Hälsingland	12	0.5%	Sweden
Medelpad	12	0.5%	Sweden
Värmland	5	0.2%	Sweden
Bohuslän	3	0.1%	Sweden

Period Distribution

Inscriptions span over 1,000 years. The major period groupings are:

Period Code	Count	%	Description
V	1,476	56.4%	Viking Age (ca. 800--1100 CE)
V (sub-periods)	384	14.7%	Viking Age with finer dating (Jelling, efter-Jelling, etc.)
M	319	12.2%	Medieval (ca. 1100--1500 CE)
M (sub-periods)	381	14.6%	Medieval with finer dating (specific centuries, years)
U	41	1.6%	Pre-Viking (Migration Period, Vendel Period, ca. 400--800 CE)
Other	14	0.5%	Uncertain dating, modern copies

Period code key: V = Viking Age, M = Medieval, U = Pre-Viking (urnordisk/Proto-Norse). Suffixes: b = beginning (början), m = middle (mitten), s = late (slutet), -t = century (talet). Example: V s 900-t = late Viking Age, 900s CE.

Data Fields

Field	Type	Description
signum	string	Unique inscription identifier from the Scandinavian Runic Text Database, e.g., "U 160", "Sö 179", "DR 41", "N 184". The prefix indicates province: U = Uppland, Sö = Södermanland, Ög = Östergötland, Öl = Öland, Vg = Västergötland, Sm = Småland, G = Gotland, Gs = Gästrikland, Vs = Västmanland, Nä = Närke, DR = Denmark, N = Norway, etc.
transliteration	string	Character-by-character rendering of the runic text using Latin letters, following Rundata scholarly conventions. Includes notation for damaged sections (...), uncertain readings ((x)), word separators (:, ·), and section markers (§A, §B)
normalization	string	Reconstruction of the text in normalized Old Norse, resolving abbreviations and damaged sections. Personal names are marked with " prefix (e.g., "Ingjaldr). Uncertain readings are in square brackets
translation	string	English translation of the inscription
period	string	Dating code from Rundata. V = Viking Age, M = Medieval, U = Pre-Viking, with optional refinements (e.g., V s 900-t = late Viking Age, 900s; M 1300-t = Medieval, 1300s)
province	string	Swedish historical province (or country for Norway/Denmark) where the inscription is located
inscription_length	int32	Number of alphabetic rune characters in the transliteration (excluding punctuation, separators, and scholarly notation)
split	string	Either "eval" (2,610 entries) or "few_shot" (5 entries selected as high-quality examples for few-shot prompting)
is_few_shot	bool	true for the 5 few-shot example entries
image	Image	Photograph of the runestone, resized to max 1024px on the longest side, saved as JPEG

Transliteration Conventions

The transliteration follows the Scandinavian Runic Text Database (Rundata) conventions:

• Lowercase Latin letters: Direct transliteration of runes (e.g., a, b, k, r, þ)
• R (capital): The ʀ rune, distinct from r in Younger Futhark
• : and ·: Word separator marks carved on the stone
• × and +: Cross-shaped separator marks
• ...: Damaged or missing section of unknown length
• -: Single missing/unreadable rune
• (x): Uncertain reading of a rune
• [text]: Scholarly reconstruction (not directly readable on the stone)
• §A, §B: Sections of the inscription (e.g., different sides of the stone)
• {text}: Inscription in a different script (e.g., Latin letters)

Example 1: Viking Age Memorial Stone (Öl 18)

Signum: Öl 18 (Öland, Sweden)

Transliteration: × inkialtr × auk × nifR × [au]k × su[i]n [× þa]iR × litu × risa ...- × iftiR × ruþmar × faþur × sin ×

Normalization: "Ingjaldr ok "Nefr ok "Sveinn, þeir létu reisa [stein] eptir "Hróðmar, fôður sinn.

Translation: Ingjaldr and Nefr and Sveinn, they had the [stone] raised in memory of Hróðmarr, their father.

Example 2: Longer Inscription (Öl 1)

Signum: Öl 1 (Öland, Sweden -- the Karlevi runestone)

Transliteration: §A + s-a... --(s)- i(a)s · satr · aiftir · si(b)(a) · kuþa · sun · fultars · in hons ·· liþi · sati · at · u · -ausa-þ-... +: fulkin : likr : hins : fulkþu : flaistr (:)· uisi · þat · maistar · taiþir : tulka · þruþar : traukr : i : þaimsi · huki · munat : raiþ:uiþur : raþa : ruk:starkr · i · tanmarku ...

Translation: This stone is set up in memory of Sibbi Góði/Goði, son of Foldarr, and his retinue set on ... Hidden lies the one whom followed (most know that) the greatest deeds, Þrud's warrior of battles, in this mound. Never will a more honest, hard-fighting 'wagon-Viðurr' upon Endill's expanses rule the land in Denmark.

Note: The Karlevi stone (Öl 1) is one of the most famous runestones in Sweden, containing the only known example of a complete skaldic stanza in runic inscription.

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("birgermoell/swedish-runestone-inscriptions")

# Access entries
entry = dataset["train"][0]
print(entry["signum"])           # e.g. "Öl 1"
print(entry["transliteration"])  # Raw runic transliteration
print(entry["translation"])      # English translation
entry["image"].show()            # Display the runestone photo

print(f"Dataset size: {len(dataset['train'])} entries")

Filtering by Split

# Get evaluation entries only (2,610 entries)
eval_entries = dataset["train"].filter(lambda x: x["split"] == "eval")

# Get few-shot examples (5 entries)
few_shot = dataset["train"].filter(lambda x: x["is_few_shot"])

Filtering by Province, Period, or Length

# Only Uppland inscriptions (the largest group)
uppland = dataset["train"].filter(lambda x: x["province"] == "Uppland")

# Only Viking Age inscriptions
viking = dataset["train"].filter(lambda x: x["period"].startswith("V"))

# Only long inscriptions (> 60 rune chars)
long_inscriptions = dataset["train"].filter(lambda x: x["inscription_length"] > 60)

# Only Norwegian inscriptions
norway = dataset["train"].filter(lambda x: x["province"] == "Norway")

Using with a Vision-Language Model

from openai import OpenAI
import base64
from io import BytesIO

client = OpenAI()
entry = dataset["train"][0]

# Convert PIL image to base64
buffer = BytesIO()
entry["image"].save(buffer, format="JPEG")
b64_image = base64.b64encode(buffer.getvalue()).decode()

response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "Transcribe the runic inscription on this stone using Latin letters. Use ':' for word separators and '...' for damaged/missing sections."},
            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{b64_image}", "detail": "high"}}
        ]
    }],
    max_tokens=1024,
    temperature=0.0,
)

prediction = response.choices[0].message.content
ground_truth = entry["transliteration"]
print(f"Prediction:   {prediction}")
print(f"Ground truth: {ground_truth}")

Evaluation Metrics

When evaluating VLM transcription performance on this dataset, we recommend:

• Character Error Rate (CER): Edit distance between predicted and ground-truth transliterations, normalized by reference length. The primary metric for this task.
• Word Error Rate (WER): Word-level edit distance, where "words" are rune sequences between separator marks.
• Normalized Levenshtein Similarity: 1 - (edit_distance / max_length), providing a similarity score in [0, 1].

All metrics should be computed at multiple normalization levels to account for the complexity of scholarly notation:

1. Lenient: Strip all scholarly notation (brackets, parentheses, section markers, ellipses), normalize separators and whitespace. Focuses purely on rune content.
2. Moderate: Preserve rune content and basic structure but normalize whitespace and separators.
3. Strict: Preserve all scholarly detail including brackets, parentheses, and separator types.

Why Multiple Normalization Levels?

Scholarly transliterations contain notation that a VLM cannot be expected to reproduce (e.g., [text] for reconstructed sections, (x) for uncertain readings). The lenient level strips all such notation to focus on what is actually carved on the stone, providing the fairest comparison. The strict level tests whether the model can also reproduce scholarly conventions.

Data Sources

Text Data

The inscription texts (transliteration, normalization, translation) come from the Scandinavian Runic Text Database (Rundata), a comprehensive database of runic inscriptions maintained by the Department of Scandinavian Languages at Uppsala University, Sweden.

• Original database: Samnordisk runtextdatabas
• Digital version used: fralik/rundata-net (SQLite export)
• The database contains ~6,500 inscriptions with transliterations; this dataset includes the 2,615 that could be matched with photographs and met quality filters.

Images

Photographs are sourced from Wikimedia Commons, matched to inscriptions via a multi-strategy search:

1. Category search: Files in Wikimedia categories like "Runestones in Uppland", "Runestones in Gotland", etc.
2. Filename search: Files whose names contain the inscription signum (e.g., "U_160", "Sö 179").
3. Full-text search: Fallback search across file descriptions.

Images were downloaded as server-generated thumbnails (960px on the longest side), then resized to max 1024px and saved as JPEG.

• Images are contributed by various photographers and are available under Creative Commons licenses (primarily CC-BY-SA)
• Each entry uses a single representative photograph of the inscription

Data Pipeline

The full pipeline code for downloading, matching, evaluating, and analyzing this dataset is available in the companion repository. The pipeline includes:

1. Download and parse Rundata SQLite database (~11,500 inscriptions)
2. Search and download images from Wikimedia Commons (3,534 images found from 6,065 candidates)
3. Join images with inscription data, apply quality filters (min 10 chars, not lost)
4. VLM evaluation with 4 prompting conditions (zero-shot, guided, few-shot-1, few-shot-3)
5. Metric computation and analysis (CER, WER, Levenshtein at 3 normalization levels)

Limitations

• Image quality variability: Photographs vary significantly in quality, lighting, angle, and stone condition. Some inscriptions may be partially obscured by lichen, moss, erosion, or poor photography.
• Damaged inscriptions: Many inscriptions are partially damaged. The transliteration includes scholarly reconstructions in brackets [text] that may not be visually verifiable from the photograph alone.
• Image-inscription mismatch: A small number of images may show a different view of the stone or site that does not clearly display the inscription. The automatic Wikimedia matching process prioritizes filename/category matches but cannot verify visual content.
• Minimum length filter: Inscriptions with fewer than 10 rune characters have been excluded, as they are typically too fragmentary for meaningful evaluation.
• Single image per inscription: Only one photograph is included per inscription, though multiple views often exist on Wikimedia Commons.
• Translation quality: English translations vary in completeness. Some are full scholarly translations, others are summaries or partial translations.
• Province imbalance: Uppland (30.1%) and Norwegian/Danish inscriptions are well-represented, while some Swedish provinces have very few entries.

Signum Prefix Reference

Prefix	Province/Region	Country
U	Uppland	Sweden
Sö	Södermanland	Sweden
Ög	Östergötland	Sweden
Öl	Öland	Sweden
Sm	Småland	Sweden
Vg	Västergötland	Sweden
G	Gotland	Sweden
Gs	Gästrikland	Sweden
Vs	Västmanland	Sweden
Nä	Närke	Sweden
HS / Hs	Hälsingland	Sweden
M	Medelpad	Sweden
Bo	Bohuslän	Sweden
Vr	Värmland	Sweden
D	Dalarna	Sweden
DR	Denmark	Denmark
N	Norway	Norway

Citation

If you use this dataset, please cite:

@dataset{swedish_runestone_inscriptions_2026,
  title={Scandinavian Runestone Inscriptions: A Multimodal Dataset for Evaluating Vision-Language Models on Ancient Runic Text},
  author={Moell, Birger},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/birgermoell/swedish-runestone-inscriptions}
}

And acknowledge the underlying data sources:

• Rundata: Samnordisk runtextdatabas, Uppsala University. https://www.nordiska.uu.se/forskn/samnord.htm
• Images: Wikimedia Commons contributors, under CC-BY-SA licenses.

License

• Dataset compilation: CC-BY-SA 4.0
• Images: Individual Wikimedia Commons licenses (predominantly CC-BY-SA 3.0 / 4.0)
• Text data: From the Scandinavian Runic Text Database (academic/research use)