Human Alignment Preferences for Text-to-Image Models, Over 1.2 Million Annotations

Over 1,200,000 human annotations form one of the largest datasets for aligning text-to-image models. Rapidata collected this data in approximately four days using their Python API, which is described as accessible for large-scale annotation. The dataset was last updated on January 10, 2025.

Use Cases

• Training reward models for text-to-image generation based on human preference annotations.
• Benchmarking the alignment of different image generation models using the collected preference data.
• Fine-tuning generative models to better match human aesthetic or safety judgments.
• Studying the distribution of human preferences across various image prompts and model outputs.

Strengths

• Contains over 1,200,000 human annotations, indicating a substantial scale.
• Data was collected rapidly in approximately four days, suggesting an efficient annotation pipeline.
• The dataset is described as one of the largest human-annotated alignment datasets for text-to-image models.

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

Rapidata

Collection Method

Collected using the Rapidata Python API for large-scale data annotation.

Freshness

Last updated 2025-01-10 22:00:00; freshness should be verified.