gpt-oss-safeguard

weights

open(gpt-oss-safeguard-20b/120b on HF)

data

not-released

license

Apache-2.0(OSI)

gpt-oss-safeguard ships open weights under Apache-2.0, but OpenAI has not released the training data or the fine-tuning pipeline behind it. That places it among models that ship permissive weights without a training recipe, rather than among fully open releases.

• https://huggingface.co/api/models/openai/gpt-oss-safeguard-120b recorded 2026-09-29

  The Hugging Face API record for openai/gpt-oss-safeguard-120b carries the tag "license:apache-2.0" and lists weight shards such as "model-00001-of-00014.safetensors", so the larger variant is published under the same license. Searched the body for dataset: none.

• https://huggingface.co/api/models/openai/gpt-oss-safeguard-20b recorded 2026-09-29

  The Hugging Face API record for openai/gpt-oss-safeguard-20b carries the tag "license:apache-2.0", is not gated, and lists downloadable weight shards such as "model-00001-of-00002.safetensors". Searched the body for dataset and datasets: none, so no training dataset is declared.

• https://huggingface.co/openai/gpt-oss-safeguard-20b recorded 2026-08-13

  Repo shows License - apache-2.0, ungated, with safetensors shards downloadable. The card documents the 20b and 120b variants and their use, and says nothing about training data - no dataset link, no mixture, and no datasets key in cardData.

• https://huggingface.co/openai/gpt-oss-safeguard-20b/raw/main/README.md recorded 2026-09-29

  The raw 20b model card front matter records "license: apache-2.0" and "base_model_relation: finetune" on openai/gpt-oss-20b, and the highlights call it a "Permissive Apache 2.0 license" that lets users "Build freely without copyleft restrictions or patent risk". Searched the card for dataset and training data: none, so no training data is published or linked.

Adoption is measured by Hugging Face downloads for the one artifact OpenAI has declared for the 20b size, openai/gpt-oss-safeguard-20b. No other channel is declared for the model, so this single download count is the only visible measure of its use.

• https://huggingface.co/api/models/openai/gpt-oss-safeguard-20b recorded 2026-10-07

  36,232 downloads in the trailing 30 days for openai/gpt-oss-safeguard-20b

gpt-oss-safeguard reasons over a policy the user writes rather than a fixed taxonomy, so it can be applied to new products and categories with minimal engineering, and it exposes its reasoning. OpenAI's own guide notes that traditional classifiers have lower latency and cost less to run.

• https://cookbook.openai.com/articles/gpt-oss-safeguard-guide recorded 2026-09-29

  The OpenAI Cookbook user guide, which says "ROOST and OpenAI have prepared a guide that explains how to write policy prompts", notes the trade-off against fixed classifiers: "Traditional classifiers have lower latency and cost less to sample from than gpt-oss-safeguard".

• https://huggingface.co/openai/gpt-oss-safeguard-20b recorded 2026-08-13

  Card describes gpt-oss-safeguard-120b and gpt-oss-safeguard-20b as 'safety reasoning models built-upon gpt-oss' that classify against a policy you supply, with a 'Bring your own policy' feature that 'interprets your written policy, so it generalizes across products and use cases', and reasoned rather than bare-label decisions.

• https://huggingface.co/openai/gpt-oss-safeguard-20b/raw/main/README.md recorded 2026-09-29

  The model card describes the pair as "safety reasoning models built-upon gpt-oss" that "classify text content based on safety policies that you provide", with a bring-your-own-policy feature that "Interprets your written policy, so it generalizes across products and use cases with minimal engineering", decisions that "Gain complete access to the model’s reasoning process", a setting to "adjust the reasoning effort (low, medium, high) based on your specific use case and latency needs", and both sizes, "gpt-oss-safeguard-120b and gpt-oss-safeguard-20b".