Guiding LLM Post-training Data Engineering with Model Internals from Sparse Autoencoders | Yi Jing et al. | ResearchPod