CORRECTION: prior draft described a joint RM+Gen objective that never ran (gen data never reached the Trainer; lambda_gen inert). Section rewritten as pure-RM training; adds matched-step LR ablation MMLU 0.5795 (1e-5) vs 0.3814 (1.5e-5) — human_annotation_results/model_training_and_mmlu.tex
8510a6b verified