Reflection details safety training and evaluation for open-source model Beam
Original title:Open model safety: lessons from training and evaluating Beam
AI summary
Reflection published a blog post detailing the safety training and evaluation methods for its open-weight model Beam and launched Project Iris, an open safety research initiative. Beam refused 89% of harmful requests in the Iris Safety Evaluations dialogue task, the highest among the open-weight models tested, while complying with 82% of safety-relevant harmless requests; after scaling the safety-alignment teacher model from 100B to 501B parameters, over-refusal fell 48.7% relative.
Key Takeaways
- Reflection published a blog post detailing the safety training and evaluation methods for its…
- Beam refused 89% of harmful requests in the Iris Safety Evaluations dialogue task,the highest…
- after scaling the safety-alignment teacher model from 100B to 501B parameters,over-refusal fell 48.
- 7% relative.
Quick Facts
- Published
- Source
- Reflection Blog
- Category
- Papers & Research
- AI score
- 66
原文拆解了 Beam 的安全训练方法与双轨评估体系,并给出关键数字,为开源模型安全实践提供可参照的路径。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Reflection details safety training and evaluation for open-source model Beam. November 6, 2026. https://new.nat6.net/en/news/mpv0jfejdzdm2336jgtyj9tde/