COCO-30-Val-2014: 30,000 Image-Caption Pairs for Generation Benchmarks

30,000 randomly sampled image-caption pairs from the COCO 2014 validation split. The dataset was created by dylanlim78 and is intended for use in image generation benchmarks. It was last updated on the platform in February 2026.

Use Cases

• Calculate FID scores based on the sampled image-caption pairs.
• Evaluate CLIPScore metrics based on the image-caption pairs.
• Benchmark image-to-text or text-to-image generation models using the structured validation split.

Strengths

• Contains 30,000 image-caption pairs, providing a substantial sample for evaluation.
• Derived from the established COCO 2014 validation split, ensuring a known provenance.
• Explicitly designed for benchmarking tasks like FID and CLIPScore.

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

Common Objects in Context (COCO) 2014 dataset.

Collection Method

Random sampling from the COCO 2014 validation split, as documented in a linked gist.

Freshness

Last updated 2026-02-01 00:05:14; freshness should be verified.