TasteBench is a benchmark of real, subjective taste judgements collected by breitburg. It contains aesthetic and creative decisions a real person made where there was no objectively correct answer, such as design direction, typography, curation, and visual style. Each item presents a situation and asks which option the person chose, with models scored against the actual human choice.
Use Cases
- Benchmarking AI models on predicting human aesthetic choices based on described scenarios.
- Training models to align with subjective human taste in creative domains like design and curation.
- Studying patterns in human decision-making where objective correctness is absent.
Strengths
- Focuses on real-world, subjective human judgements, which are often lacking in AI benchmarks.
- Items are based on actual creative decisions made by a person, providing a ground truth for evaluation.
Limitations
- Description metadata is limited; actual data quality requires manual inspection after download.
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
Provenance
- Source
- breitburg on Hugging Face
- Collection Method
- Likely collected from real human decisions on aesthetic and creative tasks.
- Freshness
- Last updated 2026-06-16 20:02:56; freshness should be verified.