DeepSeek-R1-Safe - A safe large model jointly launched by Zhejiang University and Huawei

What is DeepSeek-R1-Safe

DeepSeek-R1-Safe is a safe large model derived based onDeepSeekjointly launched by the School of Cyber Science and Technology of Zhejiang University and Huawei. The model is based on Huawei Ascend chips and the MindSpeedLLM framework. Through steps such as building safety corpora, safety supervised training, and reinforcement learning, it significantly improves the model's safety and compliance. The model has open-sourced the full-weight version and is suitable for safety training, fine-tuning, and testing, and is widely used in scenarios requiring high safety, such as cybersecurity and data protection.

Main features of DeepSeek-R1-Safe

• Safety protection function: The model can effectively identify and resist various harmful content and jailbreak attacks, with a high defense success rate, significantly improving model safety.
• General performance retention: While maintaining strong safety performance, the loss in general performance is extremely low, achieving balanced optimization of safety and performance.
• Safety training and optimization: Through technologies such as safety supervised training and reinforcement learning, the model is guided to actively identify risks and perform compliant reasoning, improving safety and robustness.
• Safety corpus construction and application: Build high-quality safety corpora and incorporate safety chains of thought, providing a solid data foundation for model training and enhancing model safety capabilities.

Technical principles of DeepSeek-R1-Safe

• Full-stack safety training framework: Starting from the underlying layer, build a full-stack safety training framework covering "high-quality safety corpora—balanced and optimized safety training—a fully self-controllable software and hardware platform across the entire chain," deeply embedding safety capabilities into the model's "thinking" and "expression."
• Safety corpus construction: By systematically reviewing 24 laws and regulations from 13 countries around the world, build a compliance benchmark covering 14 major categories of risks and achieve multi-dimensional integration of corpora. Create a triple corpus database of "risk question-safety chain of thought-safety answer," incorporating explicit safety chains of thought so that the model has active risk judgment and compliant reasoning capabilities. Introduce cutting-edge jailbreak methods to enrich attack sample strategies and guide the model to effectively resist inducement.
• Safety training paradigm: Pioneered a safety core thinking mode pre-alignment mechanism, extracting core thinking modes from safety corpora before basic training and pre-aligning them with the model's cognitive architecture to achieve rapid safety thinking guidance. Pioneered a dynamically perceived, efficient, and precise compensation mechanism, quickly compensating performance by fine-tuning non-safety-related parameters with representative data. Pioneered a multi-dimensional verifiable safety reinforcement learning mechanism, proposing a multi-dimensional fine-grained safety reward signal system, and innovatively applying a performance-safety Pareto optimal combination strategy, enabling the model to learn autonomous trade-offs and decisions in adversarial environments and achieve coordinated optimization of safety and general capabilities.

Project address of DeepSeek-R1-Safe

• GitHub repository：https://github.com/ZJUAISafety/DeepSeek-R1-Safe

Application scenarios of DeepSeek-R1-Safe

• Cybersecurity protection: The model can effectively identify and filter harmful information in the network, prevent the spread of malicious content, and protect the safety and stability of the network environment.
• Data security protection: In the process of data processing and storage, ensure the compliance and security of data and prevent data leakage and abuse.
• Content review and management: Used in content review for social media, news platforms, and others, automatically detecting and filtering violating content, improving content management efficiency.
• Intelligent customer service and dialogue systems: Provide safe and reliable content generation capabilities for intelligent customer service and dialogue systems, avoiding the generation of inappropriate or harmful responses.
• Financial risk prevention and control: In the financial sector, it is used to detect and prevent fraudulent behavior, protect the security of user funds, and maintain financial order.