You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
You agree to use this dataset for research purposes only.
Log in or Sign Up to review the conditions and access this dataset content.
DaKultur: Evaluating the Cultural Awareness of Language Models for Danish with Native Speakers
This repository covers the anonymized dataset + optional demographic information from the cultural evaluation study "DaKultur" (Müller-Eberstein et al., 2025). It includes the following fields:
input: The user input, prompting the models for knowledge of Danish culture.output_[ID]: The output generated by one of the evaluated models.judgment_[ID]: The user's judgment of the model's output (1for accept,-1for reject).comment_[ID](optional): An optional user comment regarding the model's output.topic: The human-annotated topic of the interaction.region(optional): The region where the user reports having grown up.age(optional): The user's self-reported age range in decades.gender(optional): The user's self-reported gender identity.
The evaluated models and their corresponding IDs include:
- SnakModel-7B (instruct);
- Llama2-7B (instruct) + Danish instruction-tuning;
- Llama2-7B (base) + Danish instruction-tuning.
The content of the dataset has been validated by Danish-speakers with respect to whether contents are culturally relevant, and whether judgments are coherent. For full details, please refer to the corresponding research paper.
Note that this dataset, including its demographic information are strictly for research use only, hence the non-commercial license.
- Downloads last month
- 25