Papers with Code · CC BY-SA 4.0

데이터셋 탐색

데이터셋의 설명·홈페이지를 찾고, 관련 태스크의 평가 기록을 살펴보세요.

관련 논문 14편

VisIT-Bench

VisIT-Bench 설명 보기

VisIT-Bench is a new vision-language instruction following benchmark inspired by real-world use cases. Testing 70 diverse “wish-list” skills with an automated ranking system, it advances the ongoing assessment of multimodal chatbot performance. Why VisIT-Bench 🤔? Though recent VLMs have shown promise in following instructions, their evaluation for real-world human-chatbot instructions is often limited. Typically, VLMs are evaluated through qualitative comparison of outputs, which makes it challenging to quantify progress and potential shortcomings. VisIT-Bench helps address this problem by offering a comprehensive testbed for measuring model performance across a diverse set of instruction-following tasks, inspired by real world scenarios.🌍