EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hugging Face Blog·Sep 8·5 min read
Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic
TL;DR
A new approach, Boundary-Aware Self-Distillation, allows models to refuse harmful political prompts while still answering benign ones, improving deployment safety. This method addresses the challenge of defining specific boundaries within a topic rather than rejecting the entire topic.
✦ Why It Matters
Implement Boundary-Aware Self-Distillation to enhance your model's safety in sensitive applications today.
Key Takeaways
- Implement Boundary-Aware Self-Distillation for nuanced model behavior.
- Focus on specific harmful subsets rather than broad topic refusals.
- Use coverage repair strategies to improve training data quality.
- Measure both harmful and benign prompt responses for balanced safety.
How It Works
Boundary-Aware Self-Distillation trains models to differentiate between harmful and benign prompts within the same topic. It uses a self-generation method to create training data, steering the model towards refusals on harmful prompts while retaining benign responses.
This approach includes an escalating retry strategy to ensure comprehensive coverage of difficult prompts.