quetion about scaling up.

[jinyqjinyq]

hello，I encountered a problem in my recent experiment and hope to get your answer. For the same task, after training with 100 pieces of data, the test has a relatively good success rate, but after training with 300 pieces of data, the success rate is greatly reduced. I wonder what causes this situation?

[YanjieZe]

This is interesting. If the data distribution is right, the policy usually monotonically improves. So maybe check the data first.