TASTE: Human Rankings for Design-Quality Generated Images

TASTE is a human-evaluation corpus containing panel rankings of generated images across aesthetic and faithfulness dimensions. The dataset includes per-image hallucination judgements and was released alongside the TASTE preference model. It was authored by purvanshi and last updated on Hugging Face in May 2026.

Use Cases

• Training or fine-tuning preference models based on human rankings of aesthetic quality.
• Evaluating the faithfulness of generated images to their text prompts.
• Benchmarking image generation models against human judgements for hallucination.
• Analyzing the correlation between different quality dimensions like aesthetics and faithfulness.

Strengths

• Includes human panel rankings across multiple quality dimensions, not just a single score.
• Provides per-image hallucination judgements, a specific feature mentioned in the description.
• Released alongside a named model (TASTE preference model), suggesting a defined research context.

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

Hugging Face dataset authored by 'purvanshi'.

Collection Method

Human panel rankings collected as part of the TASTE model research.

Freshness

Last updated 2026-05-29 11:32:07; freshness should be verified.