Skywork-OR1-Math-7B

A reasoning model trained in multiple stages with increasing context length, improving the performance compared to keeping the context length constant. The blog post goes into greater detail, including their entropy experiments of the actor model.