(논문 요약) Aya Vision: Expanding the worlds AI can see (Blog)

핵심 내용

• Architecture
  • 8B model: SigLIP2-patch14-384 (vision encoder) + connector + C4AI Command R7B (LLM)
  • 32B model: SigLIP2-patch14-384 (vision encoder) + connector + Aya Expanse (LLM)
• Image
  • input: 364x364
  • patch size: 28x28
  • output: 13x13
  • 364x364 보다 큰 경우 aspect-ratio 고려하여 resize
• 학습 데이터 생성
  • synthetic annotations
  • translation and rephrasing (english -> others)

성능