Self-Contradictory Instructions (SCI)
收藏资源简介:
Self-Contradictory Instructions (SCI)数据集由上海交通大学、复旦大学和上海人工智能实验室联合创建,旨在评估大型多模态模型处理自相矛盾指令的能力。该数据集包含20,000条冲突指令,均匀分布在语言和视觉范式中。通过自动数据集创建框架AutoCreate构建,该框架加速了数据集的创建过程并涵盖了广泛的指令形式。SCI数据集主要应用于提高模型对多模态指令冲突的识别能力,特别是在处理长上下文和多模态交互时。
The Self-Contradictory Instructions (SCI) dataset was co-developed by Shanghai Jiao Tong University, Fudan University, and Shanghai AI Laboratory, with the goal of evaluating the ability of large multimodal models to handle self-contradictory instructions. This dataset contains 20,000 conflicting instructions, evenly distributed across linguistic and visual modalities. It was built using the automated dataset creation framework AutoCreate, which accelerates the dataset construction process and covers a broad spectrum of instruction formats. The SCI dataset is primarily applied to improve models' capacity for identifying conflicts in multimodal instructions, particularly when processing long contexts and multimodal interactions.
数据集概述
语言-语言数据集
language-language-1
- 特征:
- context: string
- violation: string
- question: string
- 分割:
- small: 7138 bytes, 25 examples
- medium: 73709 bytes, 250 examples
- full: 831007 bytes, 2500 examples
- 下载大小: 438792 bytes
- 数据集大小: 911854 bytes
language-language-2
- 特征:
- context: string
- violation: string
- question: string
- 分割:
- small: 36214 bytes, 25 examples
- medium: 389489 bytes, 250 examples
- full: 3928775 bytes, 2500 examples
- 下载大小: 0 bytes
- 数据集大小: 4354478 bytes
language-language-3
- 特征:
- instruction1: string
- instruction2: string
- context: string
- 分割:
- small: 19597 bytes, 25 examples
- medium: 198516 bytes, 250 examples
- full: 1977170 bytes, 2500 examples
- 下载大小: 280272 bytes
- 数据集大小: 2195283 bytes
language-language-4
- 特征:
- object: string
- question: string
- prompt: string
- field: string
- 分割:
- small: 13815 bytes, 25 examples
- medium: 133962 bytes, 250 examples
- full: 1362454 bytes, 2500 examples
- 下载大小: 616010 bytes
- 数据集大小: 1510231 bytes
视觉-语言数据集
vision-language-1
- 特征:
- context: string
- img: image
- 分割:
- small: 727895.0 bytes, 15 examples
- medium: 7327050.0 bytes, 150 examples
- full: 80297026.48 bytes, 1590 examples
- 下载大小: 28095399 bytes
- 数据集大小: 88351971.48 bytes
vision-language-2
- 特征:
- context1: string
- context2: string
- img: image
- 分割:
- small: 1180429 bytes, 15 examples
- medium: 12380274 bytes, 150 examples
- full: 119183307.653 bytes, 1461 examples
- 下载大小: 123412830 bytes
- 数据集大小: 132744010.653 bytes
vision-language-3
- 特征:
- context: string
- img: image
- 分割:
- small: 196243.0 bytes, 20 examples
- medium: 1965597.0 bytes, 200 examples
- full: 19361970.0 bytes, 2000 examples
- 下载大小: 18515602 bytes
- 数据集大小: 21523810.0 bytes
vision-language-4
- 特征:
- label: int32
- question: string
- substitute_question: string
- object: string
- img: image
- 分割:
- small: 36322679 bytes, 50 examples
- medium: 224922807 bytes, 500 examples
- full: 2142965441.58 bytes, 4949 examples
- 下载大小: 453840693 bytes
- 数据集大小: 2404210927.58 bytes




