The reward on the validation set of the CALX example has been staying around 0.5 and is unable to improve or increase.

[johnson7788]

Why does the reward on the validation set of my CALX example stay around 0.5 and fail to improve? I am using a 0.5B-parameter model—could the model be too small, or is the reward too sparse? How should I improve it?

[ultmaster]

I think the performance this saturating under this configuration. You might want to tune some hyper-parameters if you want to make it better. But I doubt whether it has some real use cases since the model size is so small.