Curated mirror of openbmb/UltraData-Math maintained by the majentik MLX-quantization project.
Math-reasoning calibration/eval subset (L1 base tier) — pairs with majentik/gsm8k for reasoning-preservation checks on quantized reasoning models (MiniCPM5, Cascade-2).
4096 train samples of config UltraData-Math-L1, streamed, shuffle(seed=42). Full config set is 635 GB.… See the full description on the dataset page:
https://huggingface.co/datasets/majentik/ultradata-math-calib.