/** * @copyright Sister Software * @license AGPL-3.0 * @author Teffen Ellis, et al. * * `unit` shard recipe — US secondary-unit coverage (#451, the v0-parity `unit` gap). Onto REAL US * OpenAddresses skeletons (cached zips under `$MAILWOMAN_DATA_ROOT/oa-cache`) it INJECTS a USPS Pub-28 Appendix * C2 secondary-unit designator (the `@mailwoman/codex/us` table), varying the surface form * (canonical "Apartment" vs approved "Apt") AND the unit's POSITION (after-street / unit-first / * bare / venue-prefixed) per row, so the model learns to RECOGNIZE the designator wherever it * sits. The inline synthesis (the OA-CSV reader, the designator tables, `makeUnit`/`renderUnit`) * is ported faithfully from scripts/build-unit-shard.mjs. * * `--golden`: a held-out eval over the VERMONT source only (the corpus `defaultHoldout`, never * trained) with a different seed, emitting `{raw, components, country}` for per-locale-f1. Train * uses every NON-Vermont US source. Designators are injected in both (OA carries none), so the * eval measures designator recognition on held-out addresses. * * NOTE: this is a `generate`-mode recipe but it still reads REAL tuples off disk (`unzip` of the * cached OA zips) — `--count` bounds the OUTPUT, not the input. The passed `random` (the * framework LCG) is consumed in the exact call order the legacy script used. */ import { type ShardRecipe } from "./scaffold.ts"; /** * Shard recipe registered with the corpus builder — see the file header for the parse behaviour it exists to exercise, * and `description` below for the surface form it generates. */ export declare const unitRecipe: ShardRecipe; //# sourceMappingURL=unit.d.ts.map