Provable question dataset balance

[rcplane]

Will the benchmark saturation displayed in OpenAI's displayed evaluations inform occurrence weighting of such problem types in the provable question and answer training data that will be generated?
Which category might be beneficial to assist with sourcing and verifying data?

from
https://openai.com/index/learning-to-reason-with-llms/