Gemma: The responsible way to build
summary
Presents concrete results and tools: model comparisons show Gemma 7B is safer than Mistral 7B 58% of the time and Gemma 2B safer 56%, plus pretraining filters and safety-tuned instruction models. Releases a responsible generative AI toolkit with a parameter-efficient method and Codelab that trains a hate-speech classifier on ETHOS (800 comments) reaching state-of-the-art with only 200 labels, and an LLM prompt debugger based on saliency methods integrated with Lit.
AI-written summary · about
chapters (9)
- 0:00Introduction to responsible AI
- 0:57Google's AI principles
- 2:32Building Gemma safely
- 3:36Model evaluation strategies
- 4:46Comparative safety results
- 5:42Generative AI toolkit
- 6:50Building custom classifiers
- 7:44Debugging and prompt tuning
- 9:26Best practices for development
related talks
07mShieldGemma 2 – Developing safe and responsible AI for images
10mAI for everyone with Gemma
44mBuilding amazing Front-End Project with Gemma: A State-of-the-Art Open Model
12mWelcome to Gemma 3
39mFlutter Project with Gemma: A State-of-the-Art Open Model
09mUsing and fine-tuning Gemma 3
10mBuilding intelligent agents with Gemma 3
15mGemma, DeepMind's Family of Open Models