New Chitrakshara dataset brings multimodal AI training to 11 Indian languages with 193M images and 30B text tokens

Chitrakshara addresses the gap in Vision-Language Models (VLMs) which have been predominantly trained on English datasets, leaving Indian languages underrepresented

The dataset series includes two components: Chitrakshara-IL with 193M images, 30B text tokens, and 50M multilingual documents for pretraining, plus Chitrakshara-Cap with 44M image-text pairs containing 733M tokens

Data sourced from Common Crawl covers 11 Indian languages and includes comprehensive curation, filtering, and processing methodologies with quality and diversity analysis

Enables multi-image reasoning and understanding through large-scale interleaved image-text pretraining, moving beyond single-image focused multimodal research