Datasets:

---

danish-foundation-models

/

danish-dynaword

like 8

Follow

Danish Foundation Models 40

After examining this dataset it seems like it contains waay too many OCR errors:

Here, I plot the ratio of alphabetic characters to non-alphabetic characters:

Typically, you use a cutoff of 0.7 (Gopher paper). So this suggests that the dataset is generally low quality.

Below are some examples

the highest alpha ratio (0.79):

[...] Dagen efter Long Chriſtian der Ottendes Ded udraabtes
Kl. 9 om Morgenen Hang Majeſtæet Kong Frederik ben
Syvende til Konge fra Chriſtiansborg Slots (hvor den nye
Konge ſom Kronprinds reſiderede) Balcon af den ældfle
Statsminiſter, Geheimeraad Stemann. Den paa Pladſen for⸗
ſamlede Mængde iſtemmede Hyldingsraabet, hvorpaa Kongen
viſte ſig paa Balconen og hilſedes med Jubel. Senere toges
de paa de forſtiellige Alarmpladſe forſamlede Tropper og [...]

0.7:

[...] Suzanne.
Hor engang, han bliver toſſet⸗ ⸗ Nu da,
giv mig mit Baand igien. (hun ſöger at ſnappe b
det fra ham, men faaer ikke fat deri.)
Cherubin.
Ah! ja pyt. (han løber bag „
Suzanne. [...]

And a random one:

„^m SRat er io »or 6ommer,
^en^ Slomfler erc m;
9?aar ^r)agen ben fommcr,
(^r toor Sommer forbi.
D, ^nbigjle ^iærminbe!
D, eiffotoérofe røb!
3 aWuIm jeg fan ^Dig ftnbe,
gor ^in S)uft er jlaerf og føbl

This generally seems to suggest that the quality is not high enough to warrant when even the best ones are questionable and a large proportion of the dataset is excluded using a simple threshold.