We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Oct 29·8 min read
gpt-oss-safeguard technical report
TL;DR
Content moderation at scale requires models that can apply custom safety policies consistently, but existing systems lack transparent reasoning. OpenAI developed gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, open-weight models fine-tuned to label content against user-defined policies with explicit reasoning.
✦ Why It Matters
Engineers can deploy customizable content moderation with interpretable reasoning without retraining models for each policy change.
Key Takeaways
- Deploy gpt-oss-safeguard models to label content against custom policies without model retraining.
- Use 20b variant for resource-constrained environments; 120b for higher accuracy requirements.
- Evaluate safeguard model performance against base gpt-oss models on your specific policy domains.
- Leverage explicit reasoning outputs to audit and improve moderation decisions iteratively.
Full Summary
Content moderation systems must balance safety with policy flexibility, yet most lack interpretable decision-making. OpenAI created two open-weight safeguard models—gpt-oss-safeguard-120b (120 billion parameters) and gpt-oss-safeguard-20b (20 billion parameters)—by post-training the gpt-oss base models to reason through provided policies and classify content accordingly.
These models are designed to accept arbitrary safety policies as input and generate labeled outputs with explicit reasoning chains. The team conducted baseline safety evaluations to measure how well these safeguard variants perform relative to their underlying gpt-oss models across various content categories.
The approach enables organizations to deploy policy-agnostic moderation without retraining, though specific performance metrics and evaluation results are referenced but not detailed in this summary. This work addresses the gap between generic moderation systems and the need for transparent, policy-driven content classification at scale.