unsloth/gpt-oss-safeguard-20b
The gpt-oss-safeguard-20b is a 21 billion parameter safety reasoning model developed by OpenAI, built upon the gpt-oss architecture. It is specifically trained and tuned for classifying text content based on user-provided safety policies and performing foundational safety tasks. This model excels at LLM input-output filtering and online content labeling, providing reasoned decisions rather than just scores, and operates under a permissive Apache 2.0 license.
Loading preview...
gpt-oss-safeguard-20b: A Safety Reasoning Model
The gpt-oss-safeguard-20b is a 21 billion parameter model (with 3.6 billion active parameters) developed by OpenAI, designed for safety-critical applications. It is built on the gpt-oss architecture and is specifically trained to reason about safety policies, making it suitable for content classification and trust and safety use cases. This model is intended to fit into GPUs with 16GB of VRAM.
Key Capabilities
- Safety Reasoning: Trained and tuned to reason about safety, accommodating use cases like LLM input-output filtering and online content labeling.
- Customizable Policies: Interprets user-defined written policies, allowing for generalization across various products and use cases with minimal engineering effort.
- Transparent Decisions: Provides complete access to its reasoning process, which aids in debugging and increases trust in policy decisions. This raw Chain-of-Thought (CoT) output is intended for developers and safety practitioners.
- Configurable Effort: Users can adjust the reasoning effort (low, medium, high) to balance performance with latency requirements.
- Apache 2.0 License: Offered under a permissive Apache 2.0 license, enabling free experimentation, customization, and commercial deployment.
Usage and Integration
The model was trained on OpenAI's harmony response format and must be used with this format for correct functionality. Detailed prompting guides are available to assist in crafting policies and utilizing the model effectively. It is a partner in the Robust Open Online Safety Tools (ROOST) Model Community, reflecting a commitment to open safety and incorporating user feedback.