Loubna Ben Allal

The model actually fits on two 8xH100 https://huggingface.co/blog/open-r1/update-1#synthetic-data-generation
And the 15 generations per hour per H100 is the throughput on four nodes divided by 32 GPUs (4 to avoid the cache filling up)

We only applied Llama verification to the default subset, those rejected by Math Verify from the extended subset didn't go through a second verification step. We can release the unfiltered data with 400k problems if the community wants to do different filtering.