Source: gpt-oss-safeguard model card

Hugging Face card for OpenAI's open-weight gpt-oss-safeguard 20b/120b reasoning models that classify content against a policy supplied at inference.

Contents

Summary

gpt-oss-safeguard is a fine-tune of gpt-oss for “bring your own policy” classification: the developer writes a safety policy in natural language, the model reasons about whether content violates it and returns a label with visible reasoning. No retraining is needed to change the policy.

Key claims

  • 20b: 21B parameters, 3.6B active, fits a 16 GB GPU. 120b: 117B parameters, 5.1B active. Apache 2.0.
  • Policy is provided at inference; reasoning effort (low, medium, high) trades accuracy for latency.
  • Must be used with the harmony response format.
  • Intended for LLM input/output filtering, online content labelling and offline trust-and-safety labelling; raw reasoning is for practitioners, not end users.

Feeds

Content moderation

Suggest changes →