sullivanUCSD/anchor-400

Anchor-400是一个包含400个样本的锚定集，用于SCOUT系统，这是一个用于提示注入检测的预测器引导路由系统。锚定集是计算检测器指纹的基础：池中的每个检测器在每个锚定样本上运行一次，生成的（判决，延迟）记录被组织成指纹数据库，供SCOUT预测器在推理时检索。数据集包含230个攻击样本和170个良性样本，分为6个类别（如hidden_tricky、aligned_instruction等）和13种载体类型（如tool_output、plain_text等）。每个样本都是一个JSON对象，包含多个字段，如唯一ID、类别、载体类型、攻击类型、隐藏策略、难度标签、是否攻击、目标文本、策略文本、干净内容、评估内容、来源数据集、生成方法和备注。Anchor-400与SCOUT-450评估基准和其他外部语料库是互斥的，以确保没有检测器侧泄漏。

Anchor-400 is a 400-sample anchor set used by SCOUT, a predictor-guided routing system for prompt-injection detection. The anchor set is the substrate over which detector fingerprints are computed: every detector in the pool is run once on each anchor, and the resulting (verdict, latency) records are organized into a fingerprint database that the SCOUT predictor retrieves from at inference time. The dataset contains 230 attack samples and 170 benign samples, categorized into 6 categories (e.g., hidden_tricky, aligned_instruction) and 13 carrier types (e.g., tool_output, plain_text). Each sample is a JSON object with multiple fields, including a unique ID, category, carrier type, attack type, hiding strategy, difficulty tag, is_attack flag, goal text, policy text, clean content, eval content, source dataset, generation method, and notes. Anchor-400 is disjoint from the SCOUT-450 evaluation benchmark and other external corpora to rule out detector-side leakage.