Meet Gemma, the modern and accessible family of open models by Google inspired by the Gemini models. These large language models excel in text generation, are decoder-only, and come in English variants with open weights for further customization and specialized tuning.

Gemma 7B, a decoder-only transformer model with 7 billion parameters surpassing larger models on benchmarks, ensuring safe and responsible outputs. Google achieved this by:

• Employing automated methods to filter out personal and sensitive data from training sets
• Extensively fine-tuning through reinforcement learning from human feedback to promote responsible behaviors
• Conducting thorough evaluations, including manual red-teaming and automated adversarial testing, to mitigate risk associated with Gemma models.