Datasets:
Add instruct variant on Danish datasets (~1M words)
Early models like FLAN-T5 obtained a large degree of training data from existing structured datasets. These can be used to create reasonable instruct samples.
- dkhate (~70k)
- Euphemism-of-Death (~36k)
- lcc (~7k)
- ddisco (~100k)
- danfever (~300k)
- polstance (~5k?)
- dkstance (~100k?)
- itu_faroese_danish (~70k)
- bajer_danish_misogyny
- dacoref
- kompetencer
- danish-citizen-tests
- danish-compounds
approx sum: ~1M
An outlined examples of Euphemism-of-Death
Intended format could be:
In the following sentence is the '{target}' an euphemism for death:
{sentence_before}{sentence}{{sentence_after}
e.g.
In the following sentence is `vandrer heden` an euphemism for death:
hendes sidste levedage ; jeg formoder hun vil takke ham før hans godhed og velsigne ham før hun vandrer heden . Bed ham komme herhen , så hurtigt som muligt . Det var efter modtagelsen af dette brev , at Peter
We could probably make a variant of this (should probably be in Danish too)
estimate of size: 3 sentence of 12 words *1000 cases = 36 000
paper: https://aclanthology.org/2025.coling-main.90.pdf
github: https://github.com/mime-memo/Euphemism-of-Death
Good idea! I would leave out Danish citizen tests though, to avoid data leakage with evals.
Yeah, I did think about that as well. Nordjylland news is also in the test distribution. We could potentially move some datasets to the "test" split.
edit: Looking a bit more into it I had a few questions on the dataset (which I created a discussion on over at the repo: https://github.com/EuroEval/EuroEval/discussions/1181)