ReMem
收藏资源简介:
ReMem是由中央大学与KT公司联合构建的多模态记忆基准数据集,旨在解决大视觉语言模型(LVLM)在遗忘学习任务中基础记忆阶段的失效问题。该数据集包含100条虚构身份的全方位属性(如姓名、职业、医疗记录等),每个身份关联100组问答对(单跳与多跳问题比例7:3)及100张多样化视觉图像,通过Gemini 2.5生成文本描述并利用Nano Banana合成多视角图像。其创新性在于通过数据规模扩展、推理感知的QA架构和视觉上下文多样性,确保模型建立可靠的参数化记忆,为隐私保护中的机器遗忘研究提供严谨评估框架。
ReMem is a multimodal memory benchmark dataset jointly developed by National Central University and KT Corporation, designed to address the failure of Large Vision-Language Models (LVLMs) during the basic memory phase in forgetting learning tasks. This dataset includes 100 fictional identities with comprehensive attributes such as name, occupation, medical records, etc. Each identity is associated with 100 sets of question-answer (QA) pairs (with a 7:3 ratio of single-hop to multi-hop questions) and 100 diverse visual images. Text descriptions are generated via Gemini 2.5, and multi-view images are synthesized using Nano Banana. Its core innovation lies in ensuring that models can establish reliable parametric memories through data scale expansion, reasoning-aware QA architecture and diverse visual contexts, providing a rigorous evaluation framework for machine forgetting research in privacy protection.
数据集概述:ReMem (Reliable Multi-hop and Multi-image Memorization Benchmark)
ReMem 是一个专为大型视觉语言模型设计的可靠多跳、多图像记忆基准数据集,旨在诊断模型在隐私遗忘任务中的基础学习失败问题,确保后续遗忘评估的可靠性。
核心特性
- 发布机构:Accepted to Findings of ACL 2026(论文 arXiv:2605.03759)
- 任务领域:视觉语言模型的记忆与遗忘评估
- 数据规模:总大小约 9.5 GB,包含 4,120 个样本
- 设计理念:通过原则性数据缩放、推理感知的问答对和多样化的视觉上下文,确保模型扎实的基础学习
数据集结构
数据集包含 7 个划分,分别用于评估基础记忆阶段和遗忘阶段:
| 划分名称 | 样本数 | 描述 |
|---|---|---|
| finetune | 2,000 | 基础记忆全集,用于学习目标身份信息 |
| forget1 | 100 | 遗忘目标子集(占训练集的 5%) |
| forget2 | 200 | 遗忘目标子集(占训练集的 10%) |
| forget3 | 300 | 遗忘目标子集(占训练集的 15%) |
| forget4 | 400 | 遗忘目标子集(占训练集的 20%) |
| retain | 560 | 保留集,用于评估非目标信息的效用保持 |
| test | 560 | 测试集,用于评估遗忘鲁棒性(持出集) |
数据字段说明
每个样本包含以下字段:
| 字段名 | 类型 | 说明 |
|---|---|---|
image |
image | 个人身份图片(PIL.Image) |
question |
string | 关于敏感信息的推理感知 VQA 问题 |
answer |
string | 包含标准答案的个人信息 |
keywords |
string | 用于精确匹配评估的关键实体或值 |
question_type |
string | 推理深度(如 1-hop) |
qa_category |
string | 信息类别(如 personal_information) |
attribute |
string | 特定个人身份信息类型(如 email, date_of_birth) |
cloze_prompt |
string | 用于测量内部概率/暴露度的提示 |
image_path |
string | 原始图片文件路径 |
使用示例
通过 Hugging Face Datasets 库快速加载: python from datasets import load_dataset
加载基础记忆集
ds_full = load_dataset("herbwood27/Remem", split="finetune")
加载特定遗忘子集
ds_forget = load_dataset("herbwood27/Remem", split="forget1")
引用信息
该论文已被 ACL 2026 Findings 收录,引用格式: bibtex @article{kwon2026before, title={Before Forgetting, Learn to Remember: Revisiting Foundational Learning Failures in LVLM Unlearning Benchmarks}, author={JuneHyoung Kwon and MiHyeon Kim and Eunju Lee and JungMin Yun and Byeonggeuk Lim and YoungBin Kim}, journal={arXiv preprint arXiv:2605.03759}, year={2026} }




