hack-ignition-benchmark
收藏资源简介:
hack-ignition benchmark (v0.1) 是一个用于研究强化学习何时会产生攻击行为的数据集。它包含了在可被利用的评分器上进行强化学习(GRPO)训练的运行轨迹,每个轨迹记录了从模型、提示、训练集到评分器/奖励结构的配置信息。数据集分为三个家族(family):djinn_v2(55种配置,128次运行,79.6 MB)、mbpp(60种配置,120次运行,10.7 MB)和backdoor(28种配置,60次运行,4.1 MB)。每个家族的运行记录包含丰富的字段,如模型(Hugging Face仓库ID和版本)、提示变体、环境(评分器、奖励结构、攻击类别)、训练配方(学习率、批量大小等)、运行来源、干预调度(攻击注入、奖励切换)、系列数据(每步的攻击率、任务率和奖励率)、项目-步骤矩阵、类别序列、探针集合等。数据集故意不包含结果标签,需要用户自行设定分析视界和阈值。数据以JSON Lines格式提供,每个家族文件夹包含runs.jsonl、configs.jsonl、telemetry.jsonl等文件。该数据集适用于分析RL训练中的攻击出现时机、模式以及相关预测任务。所有运行记录和项目文件均遵循Apache-2.0许可,重分发的项目文件保留其原始来源的条款。
The hack-ignition benchmark (v0.1) is a dataset for studying when reinforcement learning induces attack behavior. It contains training trajectories from reinforcement learning (GRPO) on exploitable graders, each recording configuration information from model, prompt, training set to grader/reward structure. The dataset consists of three families: djinn_v2 (55 configurations, 128 runs, 79.6 MB), mbpp (60 configurations, 120 runs, 10.7 MB), and backdoor (28 configurations, 60 runs, 4.1 MB). Each familys runs include rich fields such as model (Hugging Face repository ID and version), prompt variant, environment (grader, reward structure, attack category), training recipe (learning rate, batch size, etc.), run provenance, intervention schedule (attack injection, reward switching), series data (attack rate, task rate, reward rate per step), project-step matrix, category sequences, probe sets, etc. The dataset intentionally omits result labels, requiring users to define their own analysis horizon and threshold. Data is provided in JSON Lines format, each family folder contains runs.jsonl, configs.jsonl, telemetry.jsonl and other files. This dataset is suitable for analyzing the timing and patterns of attack emergence in RL training and related prediction tasks. All run records and project files follow the Apache-2.0 license; redistributed project files retain terms of their original sources.




