Prepare an offline, reviewable dataset for a narrow FAQ task. Use only facts and examples I provide or explicitly authored demo records. Write JSONL with stable IDs, source versions and human review status; reject empty examples and duplicates, and keep a separate holdout set with related questions kept together. Do not call or invent a MandAPI training API. If teacher answers are later generated through ordinary Chat Completions at https://api.mandapi.com/v1, keep them as candidates and record model/prompt provenance before human review. Explain that model/provider terms must allow the intended training use. Run local JSONL validation and report dataset counts, not unmeasured training gains.