大模型安全测试问题自动生成训练数据集

目前，我国百亿级参数规模以上的大模型超过10个，十亿级参数规模以上的大模型接近80个，诸多垂直类大模型在医疗健康、教育、金融等行业中逐步开展应用。大模型的蓬勃发展也衍生出一系列的安全问题，主要包括输入型的数据投毒风险，以及输出型的违反核心价值观、输出虚假信息等风险。因此，网信部门需要能够为互联网上的各种生成式大模型产品进行备案检测和常态化的安全测评服务，从而实现对生成式大模型产品的安全治理。
针对生成式大模型安全检测需要大量测试问题，而大量测试问题需要人工参与、耗时又成本高且难以跟上大模型快速迭代的难点。本数据集可以提供给安全大模型，在使用这些数据集进行微调后，可以基于少量种子问题生成和变异出新的、大量的测试问题，从而能够解决大模型安全检测工作需要大量测试问题的难题。

Currently, there are over 10 large language models (LLMs) with a parameter scale of more than 10 billion in China, and nearly 80 LLMs with a parameter scale exceeding 1 billion. Numerous vertical-domain LLMs have gradually been deployed for applications across industries such as healthcare, education, finance and other sectors.
The rapid growth of LLMs has also spawned a series of security risks, mainly including input-side data poisoning threats, as well as output-side risks such as violations of core values and the generation of false information. Therefore, cyberspace administration authorities need to provide registration, inspection and regular security evaluation services for all generative LLM products on the Internet, so as to carry out effective security governance over generative LLM products.

A core challenge in the security detection of generative LLMs is that such detection requires a large number of test questions, which demand manual participation, are time-consuming and high-cost, and cannot keep pace with the rapid iteration of LLMs. This dataset can be provided to security-focused LLMs. After fine-tuning with this dataset, the model can generate and mutate a large quantity of new test questions based on a small number of seed questions, thereby resolving the critical challenge of requiring a massive volume of test questions for LLM security detection tasks.