据IT之家报道，知名AI安全公司DeepKeep近日发布了一份评估报告，直指Meta公司研发的Llama 2大语言模型在安全性方面存在显著问题。在涵盖13个风险评估类别的测试中，Llama 2仅通过了4项，令人对其可靠性产生质疑。
报告指出，Llama 2中的70亿参数7B模型在幻觉问题上尤为突出，即模型在生成回答时存在大量虚假或误导性内容。据DeepKeep的数据显示，该模型的幻觉率高达48%，这意味着近半数的生成回答可能包含不准确或误导的信息，这对于依赖AI模型进行决策或信息检索的用户来说，无疑是一个重大隐患。
Meta公司作为全球科技巨头，其在人工智能领域的研发备受瞩目。然而，Llama 2的这一评估结果揭示了大型语言模型在追求性能提升的同时，可能忽视了基础的安全与准确性问题。这一情况引发了业界对AI模型风险管理的深入讨论，人们呼吁在AI技术发展的同时，必须强化模型的审查和安全标准，以保障用户权益和数据安全。
DeepKeep的报告为业界敲响了警钟，Meta公司对此尚未作出正式回应。然而，这一事件无疑将促使业界对AI模型的评估和监管机制进行重新审视，以确保人工智能技术的健康发展。
英语如下：
**News Title:** “Safety Concerns Emerge Over Meta’s Llama 2 Model: DeepKeep Report Finds 48% Hallucination Rate”
**Keywords:** Meta Llama 2, safety issues, high hallucination rate
**News Content:**
**Title:** Safety Concerns Arise Over Meta’s Llama 2 Large Language Model, DeepKeep Assessment Reveals Severe Hallucination Risks
According to IT Home, renowned AI security firm DeepKeep recently released an evaluation report highlighting significant security issues with Meta’s Llama 2 large language model. In tests covering 13 risk assessment categories, Llama 2 only passed four, casting doubt on its reliability.
The report points out that the 7-billion-parameter 7B model in Llama 2 exhibits particularly striking hallucination issues, where the model generates a substantial amount of false or misleading content in its responses. Data from DeepKeep indicates that the model’s hallucination rate stands at a staggering 48%, suggesting that nearly half of its generated answers may contain inaccurate or misleading information. This poses a significant concern for users relying on AI models for decision-making or information retrieval.
As a global tech giant, Meta’s advancements in artificial intelligence have attracted considerable attention. However, Llama 2’s assessment results expose the potential trade-off between performance enhancement and fundamental safety and accuracy in large language models. The situation has sparked industry-wide discussions on AI model risk management, emphasizing the need to strengthen model scrutiny and safety standards alongside AI technology development to protect user rights and data security.
DeepKeep’s report serves as a wake-up call, with Meta yet to issue a formal response. Nonetheless, this incident is expected to prompt a reevaluation of AI model assessment and regulatory mechanisms to ensure the healthy development of artificial intelligence technology.
【来源】https://www.ithome.com/0/762/593.htm
Views: 2