VLR-BENCH
收藏资源简介:
VLR-BENCH是由首尔科学技术大学创建的一个视觉语言检索增强生成(RAG)的多语言基准数据集,旨在评估视觉语言模型(VLM)在检索增强生成任务中的能力。该数据集包含300个数据集,每个数据集包含五个不同的段落,其中只有两个段落包含解决查询所需的直接信息。数据集的内容涵盖了英语、中文和韩语的通用知识和文化数据,通过手动筛选和GPT-4o生成高质量的问题-答案对。数据集的创建过程包括图像选择、问题生成和外部知识的生成与验证。VLR-BENCH主要应用于评估模型在处理需要外部知识检索的视觉语言任务中的表现,旨在解决模型在选择有用段落以生成准确答案方面的挑战。
VLR-BENCH is a multilingual visual-language retrieval-augmented generation (RAG) benchmark dataset developed by Seoul National University of Science and Technology, designed to evaluate the performance of Vision-Language Models (VLMs) on retrieval-augmented generation tasks. This benchmark comprises 300 subsets, each containing five distinct paragraphs, with only two of the paragraphs containing the direct information required to address the corresponding query. Its content covers general knowledge and cultural data across English, Chinese, and Korean, and high-quality question-answer pairs are generated through manual screening and GPT-4o-assisted creation. The dataset creation pipeline includes image selection, question generation, as well as external knowledge generation and verification. VLR-BENCH is primarily utilized to assess model performance on visual-language tasks that require external knowledge retrieval, and it aims to resolve the challenges faced by models when selecting useful paragraphs to generate accurate answers.
数据集概述
数据集信息
-
特征(Features):
- id: 数据类型为
int64 - image_path: 数据类型为
string - instruction: 数据类型为
string - output: 数据类型为
string - input: 数据类型为
string - gold_passage_idx: 数据类型为
string - keyword1: 数据类型为
string - keyword2: 数据类型为
string - language: 数据类型为
string
- id: 数据类型为
-
数据分割(Splits):
- test: 包含 900 个样本,数据大小为 1042936 字节
-
数据集大小:
- 下载大小: 565960 字节
- 数据集大小: 1042936 字节
-
配置(Configs):
- default:
- 数据文件路径:
data/test-*
- 数据文件路径:
- default:
其他信息
- 图像下载链接: https://drive.google.com/drive/folders/1MyLwI9fWZhl733penGmCqwfUz3mQfATU?usp=sharing

- 1VLR-Bench: Multilingual Benchmark Dataset for Vision-Language Retrieval Augmented Generation首尔科学技术大学 · 2024年



