CC3M Semantic Subset: WCAG 2.2-Compliant Russian Image Captions

A curated subset of 35,794 image-caption pairs from the Conceptual Captions dataset, re-annotated in Russian for accessibility. The data was processed through semantic clustering of 2,484 groups and re-annotated using teacher vision-language models. It was created by Pavel Mikheyev and last updated in May 2026.

Use Cases

• Training Russian-language image captioning models based on WCAG 2.2-compliant descriptions.
• Evaluating model performance on semantically clustered image data.
• Benchmarking accessibility features in vision-language models.
• Fine-tuning models for specific semantic categories derived from the clustering process.

Strengths

• Contains 35,794 image-caption pairs, providing a substantial corpus.
• Captions are re-annotated in Russian to comply with WCAG 2.2 accessibility guidelines.
• Data is organized into 2,484 semantic clusters, which may aid structured training.
• Average caption length is 101 characters, suggesting detailed descriptions.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect source bias inherent to the original CC3M dataset.

Provenance

Source

Conceptual Captions (CC3M) dataset

Collection Method

Original captions were embedded, semantically clustered, and images were collected round-robin from clusters before re-annotation using teacher VLMs (Gemini, Mistral, Ollama gemma4:31b).

Freshness

Last updated 2026-05-13 03:04:59