Simple RL training for small language models

Complex RL pipelines aren't always needed for small language models. Simple training can outperform multi-stage pipelines and dynamic hyperparameter schedules.