Recently in a (github issue)[https://github.com/Kipok/NeMo-Skills/issues/214], a discussion about length-based filtering for questions came up. We had originally not done any such filtering for OpenMathInstruct-2, but found that doing such filtering doesn't affect the performance, and allows for a higher micro batch size. This PR updates the docs with this finding.
wedu-nvidia
commented
5 days ago
Recently in a (github issue)[https://github.com/Kipok/NeMo-Skills/issues/214], a discussion about length-based filtering for questions came up. We had originally not done any such filtering for OpenMathInstruct-2, but found that doing such filtering doesn't affect the performance, and allows for a higher micro batch size. This PR updates the docs with this finding.