Learn about this first
Reflection published a blog post revealing the safety training and evaluation methods for its open-weight model Beam, and launched the open-source safety research program Project Iris. Reflection said that in the Iris Safety Evaluations dialogue task, Beam refused 89% of harmful requests—the highest rate among the open-weight models tested—while complying with 82% of safety-related harmless requests.
Reflection also said that after increasing the parameters of its safety-alignment teacher model from 100B to 501B, over-refusal decreased by 48.7% relatively. All of the above percentages were results disclosed by Reflection itself.
AI-generated from reports · updated 11 days agoLatest developments
Reflection details safety training and evaluation methods for its open-source model Beam
Event progress1
Reflection releases technical report on Beam model safety training and evaluation
Reflection Blog: Reflection explains in detail the safety training and evaluation methods for its open-source model Beam