mjhermon/SecFid
收藏资源简介:
SecFid是一个基准数据集,用于衡量模型在嵌入不可信数据中的类指令内容时,安全性与保真度之间的权衡。它评估模型行为,标签包括:executed(模型遵循注入指令)、processed(模型将注入内容视为任务相关数据)、ignored(模型抑制注入内容)和other(自动检查未分类响应)。数据集包含两个配置:secfid(主文档基准,1,168行,涵盖计数、提取、翻译和编辑任务)和secfid_agentic(代理工具使用扩展,252行,源自InjecAgent风格场景)。数据集旨在评估语言模型和提示注入防御,以避免执行恶意指令,同时保留任务相关注入内容作为数据。注意:数据集包含恶意提示注入字符串,仅用于基准评估和防御研究,未经仔细过滤不应作为普通指令遵循训练数据。
SecFid is a benchmark for measuring the security-fidelity tradeoff in model behavior on instruction-like content embedded inside untrusted data. The benchmark labels model behavior into: executed (the model follows the injected instruction), processed (the model treats the injected span as task-relevant data), ignored (the model suppresses the injected span), and other (the automatic checks do not classify the response). The dataset includes two configurations: secfid (main document benchmark with 1,168 rows, covering counting, extraction, translation, and editing tasks) and secfid_agentic (agentic tool-use extension with 252 rows, derived from InjecAgent-style scenarios). It is intended for evaluating language models and prompt-injection defenses on whether they can avoid executing malicious instructions while still preserving task-relevant injected content as data. Note: The dataset contains deliberately malicious-looking prompt-injection strings and should be used for benchmark evaluation and defensive research, not as ordinary instruction-following training data without careful filtering.




