Mathematics

ProX Math CPT Evals

ProX (Programming Every Example) pre-training data refinement — released per-item math evals. Coverage is the partial slice the authors released: TinyLlama-1.1B continually pre-trained on the original OpenWebMath corpus, at two intermediate checkpoints (0.5B and 1B training tokens), scored per-item (binary correctness) on the paper's 9-task few-shot CoT math suite (GSM8K, MATH, SVAMP, ASDiv, MAWPS, TabMWP, MathQA, MMLU-STEM, SAT-Math). Per-item files for the ProX-refined CPT counterparts were not released.