Multi-turn tutoring conversations that hold one falsifiable behavioral constraint, built to instill that behavior into a small model's weights (no system prompt):

The Behavior Spec. Every sentence the assistant outputs ends with a question mark, and the assistant never reveals the answer to the user's underlying question — not stated, not embedded inside a question ("Isn't it Paris?" = fail), and not via a hint so specific it uniquely identifies the answer.… See the full description on the dataset page: https://huggingface.co/datasets/rubanikov/socratic-only-sft.