Sign in to view source links and access this dataset
Description
CulturalBench is a set of 1,227 human-written and human-verified questions designed to assess the cultural knowledge of large language models. It covers 45 global regions, including underrepresented areas like Bangladesh, Zimbabwe, and Peru. The dataset was created by author kellycyy and last updated on October 14, 2024.
Use Cases
Benchmarking LLM cultural knowledge based on the 1,227 human-verified questions.
Evaluating model performance on underrepresented regions based on coverage of 45 global areas.
Comparing model difficulty levels based on the described CulturalBench-Easy and other setups.
Strengths
1,227 human-written and human-verified questions provide a substantial evaluation corpus.
Coverage of 45 global regions, including underrepresented ones, suggests diverse cultural scope.
Human verification indicates a level of quality control for the benchmark questions.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count is unknown, which may limit suitability assessment.
Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
Source
huggingface
Collection Method
Human-written and human-verified questions.
Time Range
null
Freshness
Last updated 2024-10-14 15:27:01; freshness should be verified.
Geography
45 global regions, including Bangladesh, Zimbabwe, and Peru.
License is unknown; terms of use must be verified before application.