openai.com2025-08-05officialauthorityHN front pageabout the gpt oss familyHN 175 pts · 11 comments
cdn.openai.comofficialauthority2,803 words
OpenAI released gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, open-weight models post-trained to reason from provided policies for content classification. These text-only models achieve 46.3% multi-policy accuracy with the 120b variant and 43.6% with the 20b variant, outperforming gpt-5-thinking on internal evaluations. They support 14 languages in MMMLU evaluations and are available under the Apache 2.0 license.written, reviewed, cited
- Classifiers trained on tens of thousands of samples can outperform gpt-oss-safeguard when reasoning directly from the policy.
- The models are time and compute-intensive, making it challenging to scale across all platform content.
- The models were trained without any additional biological or cybersecurity data.
developers.openai.comofficialauthorityabout the gpt oss family
docs.unsloth.ai2025-09-27authorityHN front pageabout the gpt oss familyHN 180 pts · 43 comments1,473 words
Unsloth enables reinforcement learning training for OpenAI gpt-oss using GRPO, offering 3x faster inference and 50% lower VRAM usage compared to other implementations. Flash Attention 3 causes incorrect training losses due to lack of backward pass support for attention sinks, so Unsloth Flex Attention provides O(N) memory usage instead of O(N^2) for naive attention. Reward hacking countermeasures include inspecting imports, wiping caches, and restricting global variable access to ensure genuine task performance.written, reviewed, cited
- Unsloth achieves ~21 tokens/s for gpt-oss RL inference by rewriting code from Transformers since vLLM lacks compatibility.
- Training gpt-oss-20b with GRPO is possible on 15GB VRAM using embedding offloading which reduces usage by 1GB.
openai.com2025-10-29officialauthorityabout the gpt oss familyHN 4 pts · 1 comments
