qihoo360/TinyR1-Safety-8B
TinyR1-Safety-8B is an 8 billion parameter language model developed by qihoo360, designed for advanced content safety control in LLMs. It utilizes a unified co-training framework to integrate multiple safety behaviors, such as positive guidance, risk exposure, and refusal, which can be dynamically activated via lightweight control signals (magic tokens). This model excels at providing state-of-the-art safety alignment performance across various benchmarks, supporting flexible switching across diverse deployment scenarios and culture-aware safety policies for both English and Chinese.
Loading preview...
TinyR1-Safety-8B: Unified and Switchable LLM Safety
TinyR1-Safety-8B is an 8 billion parameter language model from qihoo360 that introduces a novel approach to content safety in large language models. Unlike traditional multi-stage safety pipelines, this model employs a unified co-training framework that integrates diverse safety behaviors within a single supervised fine-tuning (SFT) stage.
Key Capabilities & Innovations
- Dynamic Safety Control: Utilizes "magic tokens" to dynamically activate safety behaviors like positive guidance, risk exposure, and refusal, allowing for flexible switching between deployment scenarios without needing multiple specialized models.
- Culture-Aware Policies: Explores region-specific safety policies (e.g.,
policy:en-US,policy:zh-CN) via magic tokens, enabling the fusion and selective activation of diverse alignment norms for English and Chinese contexts. - High Safety Performance: Achieves state-of-the-art safety alignment, demonstrated by a Constructive Safety Score of 97.7 on its internal evaluation metric, significantly outperforming other models like Qwen3-8B (75.9) and Llama-3.1-8B-Instruct (62.1) across various safety benchmarks.
- Efficient Training: The model's design is based on data self-distillation, co-training with magic tokens for multiple safety behaviors, and safety-guaranteed generation control during inference.
When to Use This Model
- Applications requiring robust and adaptable content safety: Ideal for scenarios where LLM outputs need to be carefully moderated for safety, refusal, or constructive engagement.
- Multilingual safety contexts: Particularly useful for applications needing culture-aware safety policies in both English and Chinese.
- Developers seeking fine-grained control over safety responses: The magic token system offers unparalleled flexibility in managing how the model responds to sensitive queries.