Around 5M documents uniformly sampled from [Zyphra/dclm-dedup](https://huggingface.co/datasets/Zyphra/dclm-dedup).