Nayana & SynthDoc
Document-centric vision-language models and synthetic data across 22 languages
Overview
Nayana advances multilingual and multimodal document AI through synthetic data generation. The project includes SynthDoc, a library for generating synthetic documents, and the Nayana Synthetic Dataset, with 3 million richly annotated base documents and multilingual variants spanning 22 languages.
Key Achievements
- Meta Llama Impact Grant: Awarded $100,000 for advancing multilingual GenAI
- Publications: Papers accepted at NAACL, CVPR, and ICCV workshops
- Broad coverage: Layout, reading order, OCR, VQA, and document-retrieval annotations
SynthDoc Library
A comprehensive synthetic document generation library with:
- Support for 22 languages
- Multiple fonts and layouts
- OCR, VQA, and layout analysis capabilities
- Scalable generation using Modal with 100 CPU and 10 GPU instances in parallel
Technical Innovation
- Scalable Framework: Adapts Vision-Language Models (VLMs) for low-resource languages
- Multi-Task Support: OCR, Visual Question Answering, and layout analysis
- Production-Ready: Used to train and finetune SOTA models
Impact
The Nayana dataset addresses a critical gap in multilingual AI by providing high-quality training data for low-resource languages. This work enables more inclusive AI systems that can serve diverse linguistic communities.
Recognition
- Meta Llama Impact Grant Winner ($100,000)
- Multiple workshop acceptances at top-tier venues (NAACL, CVPR, ICCV)
- Featured in the multilingual AI research community