andrecatarino/topology-ifc-dataset

Topology-IFC数据集是为论文《Multi-Agent Memory Leakage via Topology as Information Flow Control》创建的情景记忆数据集。数据集包含来自5个Python仓库的已验证PR差异，带有污点标签和注入的私有ID令牌。数据集分为80%的训练集和20%的测试集。数据来源包括pallets/flask、psf/requests、django/django、tiangolo/fastapi和encode/httpx。记录约500个真实的PR差异，污点标签中20%为HIGH（私有），80%为LOW（公共）。HIGH记录增加了3-5个合成的私有ID令牌（PRIV-*）。训练集和测试集按80/20的比例随机分割，随机种子为42。

The Topology-IFC Dataset is an episodic memory dataset created for the paper "Multi-Agent Memory Leakage via Topology as Information Flow Control". It contains validated PR diffs from 5 Python repositories, with taint labels and injected private ID tokens. The dataset is split into 80% training and 20% test sets. Data sources include pallets/flask, psf/requests, django/django, tiangolo/fastapi, and encode/httpx. There are approximately 500 real PR diffs, with taint labels of 20% HIGH (private) and 80% LOW (public). HIGH records are augmented with 3–5 synthetic private ID tokens (PRIV-*). The train/test split is 80/20 with a random seed of 42.