ATraining's STEM RL climb — verifiable-reward training for math/physics/chemistry and competitive programming, including the multi-phase STEM data pipeline (parse → QA-pair → curate → score), competitive-coding data, and dedup/decontamination against benchmarks. Use when building verifiable STEM/math RL datasets, verifying answers (SymPy/judge/test cases), calibrating difficulty, or decontaminating training data against evals.
