Why OpenWebMath in released dataset is larger than the size reported in Readme?

[LZY-the-boys]

I noticed that the openwebmath folder in https://huggingface.co/datasets/allenai/OLMoE-mix-0924 appears to contain around 25.57B tokens, which is significantly higher than the reported 12.7B. Additionally, the arxiv folder (expected to have 21.1B tokens) seems to be missing entirely. Is the ArXiv data merged into the OpenWebMath folder, and missing its half part?

[Muennighoff]

Hmm maybe @soldni knows the answers 🤔