JMMMU
收藏资源简介:
JMMMU是由东京大学和卡内基梅隆大学共同创建的日本大规模多学科多模态理解基准,旨在评估大型多模态模型在日语环境中的表现。数据集包含1320个问题和1118张图片,涵盖28个不同学科,分为文化无关和文化特定两部分。数据集的创建过程包括翻译和文化适应,确保问题与日本文化背景相符。JMMMU主要用于评估模型在日语环境中的文化理解和语言理解能力,旨在推动多语言多模态模型的发展。
JMMMU is a large-scale Japanese multidisciplinary multimodal understanding benchmark jointly developed by the University of Tokyo and Carnegie Mellon University, aiming to evaluate the performance of large multimodal models in Japanese contexts. The dataset comprises 1,320 questions and 1,118 images, covering 28 distinct disciplines, and is split into two categories: culture-agnostic and culture-specific. The dataset creation process involves translation and cultural adaptation to ensure the questions align with Japanese cultural backgrounds. JMMMU is primarily used to assess models' cultural and linguistic comprehension capabilities in Japanese environments, with the goal of advancing the development of multilingual multimodal models.

- 1JMMMU: A Japanese Massive Multi-discipline Multimodal Understanding Benchmark for Culture-aware Evaluation东京大学 · 2024年



