The LLM-annotated corpus used to distil the ModernBERT quality scorer in Know Your Sources: Data
Selection Matters when Rewriting for Data-Constrained Pretraining.
50,427 documents, each scored by Claude Haiku 4.5 on a five-criterion rubric.
The repo name rounds to 50K. The true row count is 50,427.
OpenWebMath (mix = "openwebmath")
218… See the full description on the dataset page:
https://huggingface.co/datasets/blab-jhu/KYS-Claude-Haiku-50K-Labeled.