Housing chat distillation should teach the student to cite policy IDs the way the larger teacher does. How should losses be designed?
Select an answer to reveal the explanation.
Short Explanation
If the big housing teacher cites policy IDs, the student should learn that habit—not witty weather chat. Shape the distillation loss to transfer those citation moves.
Full Explanation
Housing chat assistants must cite municipal policy IDs the way a larger teacher does; distillation losses should be designed so the student mimics those citation-bearing rationales rather than only fluent but ungrounded prose. Transferring the teacher’s municipal citation habits via the distillation objective improves auditability and resident trust in eligibility and rebate guidance, which housing agencies need for appeals and oversight. Penalizing any citation so the student never mentions policy IDs actively suppresses the grounding behavior housing agencies need and increases unsupported advice risk. Distilling only punctuation timing and ignoring rationales with citations transfers superficial form, not policy references residents can verify. Training the student solely on unrelated weather tweets as soft targets moves the student away from housing policy space entirely and wastes the teacher signal. Exam caveat: citation mimicry is not verification—pair distilled citation style with retrieval or allow-list checks so fabricated policy IDs cannot ship just because they look teacher-like. Operational check: score citation ID exact-match and unsupported-citation rate on a housing gold set for teacher versus student, and require an allow-list validator in the serving path before resident cutover.