Performance on Out-of-Distribution Datasets

[biandh]

Have you tried evaluating the performance of the model after RL on datasets that are out of the distribution of the RL training data?
Is there still a noticeable improvement ?

[michaelzhiluo]

We have some more evaluation on Olympiad and Minera:
We do show improvement over Deepseek's original distilled model, it was not that good on Minerva but we crush others on OlympiadBench.