Sign in to view source links and access this dataset
Description
GlobeAudio is a benchmark for evaluating large audio-language models, comprising 5,637 human-authored multiple-choice questions. The dataset covers six typologically diverse languages, including English, Chinese, Thai, and Russian. It was created by iNLP-Lab and last updated in June 2026.
Use Cases
Benchmarking model performance on naturalistic audio understanding based on the described multiple-choice questions.
Evaluating cross-lingual and cross-cultural generalization of audio-language models based on the six-language coverage.
Training models for audio question-answering tasks based on the human-authored and verified MCQs.
Strengths
Contains 5,637 human-authored and verified multiple-choice questions.
Covers six typologically diverse languages, including English, Chinese, Thai, and Russian.
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
iNLP-Lab
Collection Method
Human-authored and rigorously verified, as described for the associated research paper.
Freshness
Last updated 2026-06-09 16:01:15; freshness should be verified.
Geography
Multicultural, with languages from the United States, China, Thailand, and Russia.
License is unknown; terms of use must be verified before application.