A quality-controlled human preference dataset for text-to-image generation. It contains 40,000 trust-weighted pairwise judgments from calibrated annotators, comparing AI-generated images on prompt alignment and overall preference. This subset, created by datapointai, is described as the highest-annotator-quality version.
Use Cases
- Fine-tuning text-to-image models based on human preference data for prompt alignment.
- Training reward models for reinforcement learning from human feedback (RLHF) based on pairwise comparisons.
- Benchmarking image generation models based on human judgments for overall preference.
- Studying the correlation between prompt alignment and overall aesthetic preference in generated images.
Strengths
- Contains 40,000 trust-weighted pairwise human judgments.
- Data is sourced from calibrated annotators, suggesting a focus on annotation quality.
- Judgments are made across two distinct evaluation dimensions: prompt alignment and overall preference.
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
- datapointai via the Hugging Face platform.
- Collection Method
- Built on the Datapoint annotation platform using calibrated human annotators.
- Freshness
- Last updated 2026-03-30 06:37:56; freshness should be verified.