Final yr, we launched ShieldGemma, a collection of security content material classifier fashions constructed on Gemma 2 and designed to detect dangerous content material in AI fashions’ textual content inputs and outputs. As we debut Gemma 3 right this moment, we’re excited to construct on our basis of accountable AI by asserting ShieldGemma 2.
ShieldGemma 2, constructed on Gemma 3, is a 4 billion (4B) parameter mannequin that checks the security of your artificial and pure pictures in opposition to key classes that can assist you construct strong datasets and fashions. With this addition to the Gemma household of fashions, researchers and builders can now simply reduce the chance of dangerous content material of their fashions throughout key areas of hurt:
- Sexually specific content material
We advocate utilizing ShieldGemma 2 as an enter filter to imaginative and prescient language fashions, or as an output filter of picture era programs. ShieldGemma can be utilized on each artificial and pure pictures.
What’s completely different in ShieldGemma 2?
Transferring past textual content, coaching and understanding picture security in multimodal fashions brings new challenges, which is why ShieldGemma 2 is constructed to reply to a variety of numerous and nuanced types of images.
To coach a sturdy picture security mannequin, we curated coaching datasets of pure and artificial pictures, and instruction-tuned Gemma 3 to display sturdy efficiency. We in contrast security insurance policies to the next benchmarks, and will likely be releasing a technical report that additionally incorporates third social gathering benchmarks.
Analysis outcomes primarily based on optimum F1 rating (%, greater is best) on our inside benchmark
Right here’s how ShieldGemma will help you construct safer AI picture purposes:
- Flexibility: Add any artificial or pure pictures, and edit our immediate template to adapt to your wants. Fantastic-tune on Google Colab or your personal GPU.
- Versatility: All instruments that assist Gemma 3 assist ShieldGemma 2, together with common frameworks like Transformers, JAX, Keras, Ollama, and others.
- Collaborative: ShieldGemma is open by nature and welcomes group collaborators to maintain constructing inclusively as we collectively push business security requirements onwards and upwards.
Deploying open fashions responsibly depends on a complete group effort, and we sit up for exploring how ShieldGemma 2 may be delivered in smaller sizes, throughout extra hurt areas, and aligned with multimodal ML Commons taxonomy within the close to future.
We’re excited to proceed constructing for protected and accountable multimodal AI!
Get began right this moment
- Discover ShieldGemma 2 on our developer website, and see our mannequin card for extra info.
- Attempt ShieldGemma 2 on Google AI Studio, Hugging Face, Ollama, and different platforms.
Crew Acknowledgement
Wenjun Zeng, Ryan Mullins, Dana Kurniawan, Yuchi Liu, Mani Malek, Yiwen Tune, Dirichi Ike-Njoku, Hamid Palangi, Jindong Gu, Shravan Dheep, Karthik Narashimhan, Tamoghna Saha, Joon Baek, Rick Pereira, Cai Xu, Jingjing Zhou, Aparna Joshi, Will Hawkins
