ISACL dataset
收藏资源简介:
ISACL数据集是专门为训练神经网络分类器以识别版权或专有数据泄露风险而精心策划的。数据集由版权材料及其基于相似性评分的风险标签组成,这些相似性评分是从LLM在预填充阶段的内部状态中提取的。该数据集用于训练模型,以便在内容生成之前主动评估潜在的风险,从而实现早期干预并防止敏感信息的暴露。
The ISACL dataset is meticulously curated specifically for training neural network classifiers to identify copyright or proprietary data leakage risks. The dataset consists of copyrighted materials and their risk labels based on similarity scores extracted from the internal states of LLMs during the prefill stage. This dataset is used to train models to proactively evaluate potential risks prior to content generation, enabling early intervention and preventing the exposure of sensitive information.
数据集概述
数据集名称
Internal states leakage
数据集来源
通过执行"git clone https://github.com/chentong0/copy-bench.git"命令获取,位于"copyright"文件夹下
数据处理流程
- 生成输出(scripts/generate.py)
- 评估rouge分数(scripts/eval_literal_copying.py)
- 划分为侵权/非侵权类别(scripts/label.py)
- 通过带标签的输入提取内部状态,使用内部状态训练customMLP(scripts/train.py --> ipynb)
- 1ISACL: Internal State Analyzer for Copyrighted Training Data LeakageCity University of Hong Kong, Microsoft, University of Illinois Urbana-Champaign, Google LLC, Arizona State University, Stevens Institute of Technology · 2025年



