ShieldGemma

weights

open(shieldgemma 2B/9B/27B on HF)

data

not-released

license

Gemma(NOT OSI, use-restricted)

Google publishes the weights under the Gemma license, which carries use restrictions and is not OSI-approved. The fine-tuning data, described as synthetic internal data plus public datasets, is not released.

• https://huggingface.co/api/models/google/shieldgemma-2b recorded 2026-09-29

  The Hugging Face API record for google/shieldgemma-2b carries the tag "license:gemma", is gated for manual approval, and lists the weights as "model-00001-of-00002.safetensors" and "model-00002-of-00002.safetensors"; its gating prompt says "you’re required to review and agree to Google’s usage license" and its cardData declares no datasets.

• https://huggingface.co/google/shieldgemma-2b recorded 2026-09-29

  The model card describes the models as "available in English with open weights, including models of 3 sizes: 2B, 9B and 27B parameters" and says only that they were "fine-tuned on synthetically generated internal data and" public datasets, linking no training data release.

Adoption is measured by Hugging Face downloads for google/shieldgemma-2b, one of three published sizes. Most use of the model likely happens through Google's Gemma tooling rather than through direct downloads of this checkpoint, so the figure counts only part of its use.

• https://huggingface.co/api/models/google/shieldgemma-2b recorded 2026-10-07

  12,594 downloads in the trailing 30 days for google/shieldgemma-2b

ShieldGemma screens prompts and responses against four harm categories in English, in 2B, 9B and 27B sizes. Llama Guard covers fourteen categories in eight languages, so ShieldGemma's narrower coverage sits below it.

• https://huggingface.co/google/shieldgemma-2b recorded 2026-09-29

  The model card says ShieldGemma is "a series of safety content moderation models built upon" Gemma 2 "that target four harm categories (sexually explicit, dangerous content, hate, and harassment)", and its chat template asks whether "a user prompt is in violation of the defined safety policies" or whether "an AI response to a prompt is in violation of the defined safety policies", answering "Your answer must start with 'Yes' or 'No'".