ShieldGemma model card

Model Page: ShieldGemma

Resources and Technical Documentation:

• Responsible Generative AI Toolkit
• ShieldGemma on Kaggle
• ShieldGemma on Hugging Face Hub

Terms of Use: Terms

Authors: Google

Model Information

ShieldGemma 2 is a model trained on Gemma 3’s 4B IT checkpoint for image safety classification across key categories that takes in images and outputs safety labels per policy.

Description

ShieldGemma 2, built on Gemma 3, is a 4 billion (4B) parameter model that checks the safety of both synthetic and natural images against key categories to help you build robust datasets and models. With this addition to the Gemma family of models, researchers and developers can now easily minimize the risk of harmful content in their models across key areas of harm as defined in the guidelines below.

We recommend using ShieldGemma 2 as an input filter to vision language models, or as an output filter of image generation systems. To train a robust image safety model, we curated training datasets of natural and synthetic images and instruction-tuned Gemma 3 to demonstrate strong performance. We compared safety policies to different benchmarks, and are releasing a technical report that incorporates third party benchmarks.

Inputs and outputs

• Input: Image and prompt instruction with example policy definition above.
• Output: Probability of 'Yes' or 'No' tokens, with a higher score indicating the model's higher confidence that the image violates the specified policy. 'Yes' means that the image violated the policy, 'No' means that the model did not violate the policy.

Guidelines to use in prompts

Here is our prompt instruction [Prompt Instruction]:

You are a policy expert trying to help determine whether a generated or real image is in violation of the safety policy below. Your decision must very clearly follow the specific language used in the safety policy and should not include any interpretation outside of the safety policy.

<BEGIN_SAFETY_POLICY>
{policy}
</END_SAFETY_POLICY>

Does the image violate the above policy? Your Answer must start with 'Yes' or 'No'.

Use Case: Image Classification

Citation

@article{shieldgemma2,
    title={ShieldGemma 2},
    url={https://ai.google.dev/gemma/docs/shieldgemma/model_card_2},
    author={ShieldGemma Team},
    year={2025}
}

Model Data

Data used for model training and how the data was processed.

Training Dataset

Our training dataset consists of both natural images and synthetic images. For natural images, we sample a subset of images from WebLI (Web Language and Image) dataset that are relevant to the safety tasks. For synthetic images, we leverage an internal data generation pipeline to enable controlled generation of prompts and corresponding images that balance the diversity and severity of images. For this study, harm types were limited to dangerous, sexually explicit, and violent content, with English only. Additional adversarial and sub-topics were structured using a taxonomy that corresponds to respective policies, and a range of demographics, context, and regional aspects.

Data Preprocessing

Here are the key data cleaning and filtering methods applied to the training data: CSAM Filtering: CSAM (Child Sexual Abuse Material) filtering was applied in the data preparation process to ensure the exclusion of illegal content.

Implementation Information

Hardware

ShieldGemma 2 was trained using the latest generation of Tensor Processing Unit (TPU) hardware (TPUv5e), for more details refer to the Gemma 3 model card.

Software

Training was done using JAX and ML Pathways. For more details refer to the Gemma 3 model card.

Evaluation

Benchmark Results

ShieldGemma 2 4B was evaluated against internal and external datasets. Our internal dataset is synthetically generated through our internal image data curation pipeline. This pipeline includes key steps such as problem definition, safety taxonomy generation, image query generation, image generation, attribute analysis, label quality validation, and more. We have approximately 500 examples for each harm policy. The positive ratios are 39%, 67%, 32% for sexual, dangerous content, violence respectively. We will also be releasing a technical report that includes evaluations against external datasets.

Internal Benchmark Evaluation Results

Ethics and Safety

Evaluation Approach

Although the ShieldGemma models are generative models, they are designed to be run in scoring mode to predict the probability that the next token would Yes or No. Therefore, safety evaluation focused primarily on outputting effective image safety labels.

Evaluation Results

These models were assessed for ethics, safety, and fairness considerations and met internal guidelines. When compared with benchmarks, evaluation datasets were iterated on and balanced against diverse taxonomies. Image safety labels were also human-labelled and checked for use cases that eluded the model, enabling us to improve upon rounds of evaluation.

Usage and Limitations

These models have certain limitations that users should be aware of.

Intended Usage

ShieldGemma 2 is intended to be used as a safety content moderator, either for human user inputs, model outputs, or both. These models are part of the Responsible Generative AI Toolkit, which is a set of recommendations, tools, datasets and models aimed to improve the safety of AI applications as part of the Gemma ecosystem.

Limitations

All the usual limitations for large language models apply, see the Gemma 3 model card for more details. Additionally, there are limited benchmarks that can be used to evaluate content moderation so the training and evaluation data might not be representative of real-world scenarios.

ShieldGemma 2 is also highly sensitive to the specific user-provided description of safety principles, and might perform unpredictably under conditions that require a good understanding of language ambiguity and nuance.

As with other models that are part of the Gemma ecosystem, ShieldGemma is subject to Google's prohibited use policies.

Ethical Considerations and Risks

The development of large language models (LLMs) raises several ethical concerns. We have carefully considered multiple aspects in the development of these models.

Refer to the Gemma 3 model card for more details.

Benefits

At the time of release, this family of models provides high-performance open large language model implementations designed from the ground up for Responsible AI development compared to similarly sized models.

Using the benchmark evaluation metrics described in this document, these models have been shown to provide superior performance to other, comparably-sized open model alternatives.