Reflection details safety training and evaluation methods for its open-source model Beam
Open model safety: lessons from training and evaluating Beam
AI BriefingReflection published a blog post explaining the safety training and evaluation methods for its open-weight model Beam and launched Project Iris, an open-source safety research initiative. Beam refused 89% of harmful requests in Iris Safety Evaluations dialogue tasks, the highest rate among the open-weight models tested, while complying with 82% of benign safety-related requests. Scaling the safety alignment teacher model from 100B to 501B parameters reduced over-refusal by 48.7% relative to the baseline.
Source: Reflection Blog · reflection.ai
Event follow-up1
View all event coverage- Reflection details safety training and evaluation methods for its open-source model BeamReflection Blog · 11 days ago