little-learner

This article presents LittleLearner, a controlled sandbox for studying LLM knowledge acquisition using a K-5 curriculum-filtered dataset, finding that interventions like scaling and post-training enhance in-scope performance but do not improve out-of-scope capabilities.