ADS-Edit
收藏资源简介:
ADS-Edit是一个专为自动驾驶系统设计的多模态知识编辑数据集,由浙江大学的研究团队创建。该数据集包含多种真实世界场景,涵盖视频、多视角图像和单张图像三种数据类型,旨在评估模型在自动驾驶场景下的知识编辑能力。数据集的构建经历了数据预处理、可靠性数据构建、泛化数据构建和局部性数据构建等步骤,并采用自动化流程结合人工审核确保数据质量。该数据集可应用于自动驾驶领域,解决大型多模态模型在直接应用于自动驾驶系统时面临的挑战。
ADS-Edit is a multimodal knowledge editing dataset specifically designed for autonomous driving systems, created by the research team from Zhejiang University. This dataset covers a variety of real-world scenarios and includes three data types: video, multi-view images, and single images, aiming to evaluate the knowledge editing capabilities of models in autonomous driving scenarios. The construction of the dataset involves several steps including data preprocessing, reliability data construction, generalization data construction, and locality data construction, and adopts an automated workflow combined with manual review to ensure data quality. This dataset can be applied in the field of autonomous driving to address the challenges faced when large multimodal models are directly applied to autonomous driving systems.
EasyEdit 数据集概述
数据集简介
EasyEdit 是一个用于大型语言模型(LLM)知识编辑的框架,旨在通过特定的编辑技术调整模型的行为,使其在特定领域内的表现得到优化,同时不影响其他输入的性能。
数据集内容
编辑任务定义
- 知识插入:向模型中注入新的知识。
- 知识更新:更新模型中已有的过时知识。
- 知识擦除:擦除模型中的敏感信息。
编辑技术比较
EasyEdit 支持多种知识编辑技术,包括:
- Memory-based: SERAC, IKE, GRACE, MELO, WISE
- Meta-learning: MEND, InstructEdit, MALMEN
- Locate-then-edit: KN, ROME, MEMIT, PMET, DINM, R-ROME, EMMET
- FT-L
评估指标
- Reliability: 编辑成功率
- Generalization: 编辑范围内的成功率
- Locality: 编辑后对无关输入的影响
- Portability: 编辑在推理/应用中的成功率
- Efficiency: 时间和内存消耗
数据集版本
- 版本: v0.0.1
- 最新提交: May
数据集链接
- 论文: A Comprehensive Study of Knowledge Editing for Large Language Models
- 基准: KnowEdit
- 代码: EasyEdit GitHub
相关资源
- 文档: EasyEdit Doc
- 演示: EasyEdit Demo
- 教程: EasyEdit Slides
- 视频: EasyEdit Video

- 1Benchmarking Chinese Knowledge Rectification in Large Language Models浙江大学 · 2024年



