mukunda1729/jailbreak-corpus-mini

Jailbreak Corpus Mini是一个小型的手工收集的jailbreak和prompt-injection测试用例集合，专门用于测试大型语言模型（LLM）应用的防护机制。数据集包含15个测试用例，分布在8个类别（如prompt_injection、prompt_leak等）和3个严重性级别（高、中、低）中。每个测试用例都是一个JSON对象，包含稳定的标识符（id）、对抗性输入字符串（prompt）、类别（category）和严重性（severity）字段。该数据集旨在帮助开发者在测试套件中捕获回归问题，为新团队成员提供一个自包含的对抗性尝试库，并为防护工具提供一个可重复的测试用例集合进行评估。需要注意的是，该数据集不是一个基准测试，也不适用于生产模型的训练。

Jailbreak Corpus Mini is a small, hand-curated set of jailbreak and prompt-injection fixtures for testing LLM-app guardrails. It includes 15 fixtures spread across 8 categories (e.g., prompt_injection, prompt_leak) and 3 severity levels (high, medium, low). Each fixture is a JSON object containing a stable identifier (id), an adversarial input string (prompt), a category (category), and a severity level (severity). The dataset is designed to help developers catch regressions in their test suites, provide new team members with a self-contained library of adversarial attempts, and offer a reproducible fixture set for evaluating guardrail tooling. Note that this is not a benchmark and is not safe for production model training.