@mailwoman/corpus 9.1.0 → 9.3.0

package/lib/recipes/scaffold.ts

@@ -0,0 +1,469 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Shared scaffolding for the synthetic-corpus SLICE RECIPES — the common bits the 16
+ *   `build-*-slice.mjs` scripts each re-implemented: the seeded LCG PRNG, the tuple reader, and the
+ *   canonical → `alignRow` → `LabeledRow` JSONL emit step. A recipe ({@link CorpusRecipe}) supplies
+ *   only its synthesis + filter; the `mailwoman corpus slice <recipe>` command supplies the I/O.
+ */
+
+import { pathExists } from "@mailwoman/core/fs/readers"
+import { readZipEntry } from "@mailwoman/core/fs/zip"
+import { tryParsingJSON } from "@mailwoman/core/json"
+import type { PathBuilderLike } from "path-ts"
+import type { AsyncChunkIterator, AsyncDataResource } from "spliterator"
+import { AsyncSequence, CSVSpliterator, TextSpliterator } from "spliterator"
+
+import { stableSourceIDFromParts } from "#adapters/utils"
+import { alignRow } from "#utils"
+
+/**
+ * {@link stableSourceIDFromParts} under the name the recipes use: arbitrary disambiguator keys (e.g. a variant index
+ * `v`) that aren't `ComponentTag`s, which is how the legacy builders kept per-variant ids unique.
+ */
+export function sliceSourceID(adapterID: string, parts: Record<string, string | undefined>): string {
+	return stableSourceIDFromParts(adapterID, parts)
+}
+
+/**
+ * Where a country's convention writes the postcode inside the `«locality», «region»[, «country»]` admin tail.
+ *
+ * Position changes the tag assigned to the same digits. On the shipped model, `Heladería Frappé Manía, Avenida Country
+ * Club, Barcelona 6001, Anzoátegui, Venezuela` tags `6001` as `house_number` and loses the locality into the street,
+ * while the identical row written `… 6001 Barcelona, Anzoátegui, Venezuela` tags it `postcode` and recovers `locality:
+ * Barcelona`. So a slice emitting one placement teaches one family of countries.
+ *
+ * Each is attested by a gauntlet board row, which is the bar for adding another:
+ *
+ * - `leading` — `«postcode» «locality», «region»`. `Rua da Praia, 15, 8600-315 Lagos, Algarve, Portugal`
+ *   (`pt_structured`). The default, and what every tuple written before this field existed means.
+ * - `after_locality` — `«locality» «postcode», «region»`. `…, Barcelona 6001, Anzoátegui, Venezuela`
+ *   (`ve_city_postcode_trailing_state`).
+ * - `after_region` — `«locality», «region» «postcode»`. `12 MG Road, Indiranagar, Bengaluru, Karnataka 560038, India`
+ *   (`in_structured`).
+ */
+export type PostcodePlacement = "leading" | "after_locality" | "after_region"
+
+/**
+ * A (locality, region, postcode, country) source tuple — the input to tuples-mode recipes.
+ */
+export interface SliceTuple {
+	locality?: string
+	/**
+	 * The segment before the locality, when the source has one. A slice whose every row BEGINS with the locality teaches
+	 * that the first named segment is the locality, which flips the model's default — measured on the v4.8.0 candidate,
+	 * where `Ye Three Lords, 27 Minories, London EC3N 1DE` came back `locality: "Ye Three Lords"`.
+	 */
+	dependentLocality?: string
+	region?: string
+	postcode?: string
+	country?: string
+	/**
+	 * Defaults to `leading` when absent, which is what every tuple file written before this field existed means — so an
+	 * old tuples file produces byte-identical rows.
+	 */
+	postcodePlacement?: PostcodePlacement
+	[k: string]: unknown
+}
+
+/**
+ * The two seeded generators the legacy `build-*-slice` scripts used, re-exported from their home in
+ * `@mailwoman/core/utils` so a recipe keeps importing everything it needs from this one scaffold module.
+ *
+ * - `makeLcg` (`s = s*1664525 + 1013904223 mod 2^32`) — what the street/po-box/anchor builders seeded.
+ * - `makeMulberry32` — what the MAJORITY of them used (german, locale, boundary-stress, unit, fr-order, country-balanced,
+ *   intersection, fr-admin-split, street-affix, street-bare, po-box-cedex).
+ *
+ * A recipe must seed the same one its `.mjs` did, the same way it did (usually `seed`, but some derive a per-stream
+ * seed), or `--seed N` stops being byte-reproducible.
+ */
+export { makeLcg, mulberry32 as makeMulberry32, makeLcg as makeRandom } from "@mailwoman/core/utils"
+
+/**
+ * One CSV record, keyed by its lower-cased header name, every value trimmed. A column the header does not declare reads
+ * as `undefined`; a column the header declares but this record does not reach reads as `""`.
+ */
+export type CSVRecord = Record<string, string | undefined>
+
+/**
+ * Quote handling spans the ROW split, not only the column split: a newline inside a quoted field belongs to a single
+ * record, so the record boundaries can only be found by a scanner that already knows where the quotes are. Find the
+ * lines first and unquote afterwards and such a record splits in two — the first half short by however many columns
+ * followed the newline, which is how a `street` value comes to be read as `city`.
+ *
+ * `header: false` keeps the first record in the stream, so this module owns the lower-casing rather than
+ * `normalizeKeys` — see {@link toHeader}.
+ */
+const CSV_INIT = { header: false, enableQuoteHandling: true } as const
+
+/**
+ * Header names are lower-cased on the way in. `normalizeKeys` will not do it: it leaves an ALL CAPS header alone, and
+ * OpenAddresses ships `LON,LAT,NUMBER,STREET` while other extracts ship the same names lower-case. A recipe names its
+ * columns in lower case either way.
+ */
+function toHeader(cells: readonly string[]): string[] {
+	return cells.map((name) => name.trim().toLowerCase())
+}
+
+/**
+ * Line breaks inside a value become single spaces. A quote-aware parse is the first thing here able to return a value
+ * CONTAINING one — `us/ia/statewide.csv` has 12, all unit designators like `"#2\n#2"` — and every consumer synthesizes
+ * one-line address text from these cells with no guard, because until that parse landed no value could carry one.
+ * Collapsing keeps the record (the address is fine; the source's line break is not part of it) without emitting a
+ * training row with a newline inside it.
+ *
+ * Only `\r` and `\n`, deliberately — NOT `\s`. Runs of spaces and tabs pass through exactly as the source wrote them
+ * (OA's IA extract writes `NORTH`, three spaces, `MAIN STREET`), because those could always appear and every slice
+ * built to date contains them. Widening this to `\s+` silently rewrites values on rows with no line break at all.
+ * `scaffold.test.ts` pins both halves, against both readers.
+ */
+function toRecord(header: readonly string[], cells: readonly string[]): CSVRecord {
+	const record: CSVRecord = {}
+
+	for (let i = 0; i < header.length; i++) {
+		record[header[i]!] = (cells[i] ?? "").replaceAll(/[\r\n]+/g, " ").trim()
+	}
+
+	return record
+}
+
+/**
+ * Read a CSV as header-keyed records.
+ *
+ * Returns the spliterator's own {@linkcode AsyncSequence}, so a caller composes `take`/`drop`/`filter` onto it — those
+ * ops fuse into one pull loop, and a `take` that is satisfied closes the source's file handle on the way out. Wrapping
+ * this in an `async function*` costs an async frame per row and takes those ops away; don't.
+ *
+ * A source at or below the spliterator's 128 KiB bulk threshold is read whole and parsed by the synchronous engine, so
+ * this is also the right reader for small sources — there is no buffered variant to reach for.
+ *
+ * @category CSV
+ */
+export function readCSVRecords(source: AsyncDataResource | AsyncChunkIterator): AsyncSequence<CSVRecord> {
+	let header: string[] | null = null
+
+	return CSVSpliterator.fromAsync<string[]>(source, CSV_INIT)
+		.filter((cells) => {
+			if (header) return true
+
+			header = toHeader(cells)
+
+			return false
+		})
+		.map((cells) => toRecord(header!, cells))
+}
+
+/**
+ * {@link readCSVRecords} over one member of a zip archive — what every recipe reading a cached OA source wants.
+ *
+ * A source a checkout has not cached yields nothing, after saying so. A lab holds the archives for the countries it has
+ * built, so a recipe naming ten sources routinely finds three, and the `unzip -p` subprocesses these replaced behaved
+ * the same way by accident — a non-zero exit warned and returned no rows. A recipe that ends up with NO tuples at all
+ * still throws; that is the case where the cache, not the recipe, is the problem.
+ *
+ * @category CSV
+ */
+export function readZippedCSVRecords(archivePath: PathBuilderLike, entryName: string): AsyncSequence<CSVRecord> {
+	return AsyncSequence.from<CSVRecord>(async () => {
+		if (!(await pathExists(String(archivePath)))) {
+			console.error(`  WARN: ${archivePath} is not cached — skipping ${entryName}`)
+
+			return []
+		}
+
+		return readCSVRecords(readZipEntry(archivePath, entryName))
+	})
+}
+
+/**
+ * License stamped on the synthetic tuple-derived slices (`po-box`, `no-street`, `house-venue`): the output is
+ * generated, but it inherits the terms of the real tuples it is derived from, so the attribution travels with it.
+ */
+export const SYNTHETIC_TUPLE_LICENSE = "Synthetic — derived from CC-BY / public-domain input tuples"
+
+/**
+ * The surface key shared by the Norwegian slices (`no-fragment`, `no-street-led`).
+ *
+ * MUST match the NO digit board's `norm_surface`: NFC, lowercase, collapse whitespace — and KEEP diacritics.
+ * fr-fragment's norm strips them (NFD + combining-mark removal), which is right for French but would collapse
+ * `Tømmerlien` → `tommerlien` here, so a slice's exclusion check would never match the board's reserved `tømmerlien`
+ * and the train/eval split would leak silently. Diacritic street heads (…vegen/…veien with ø/å/æ) are the whole point
+ * of those slices' boundary; folding them away is not an option.
+ */
+export const foldNOSurface = (value: string): string =>
+	value.normalize("NFC").toLowerCase().replaceAll(/\s+/g, " ").trim()
+
+/**
+ * A cached OpenAddresses extract: the zip and the CSV member.
+ */
+export interface OATupleSource {
+	zip: PathBuilderLike
+	csv: string
+}
+
+/**
+ * The four base fields every OA tuple reader extracts. `postcode` is `""` when the row carries none and
+ * {@link ReadOATuplesOptions.requirePostcode} is unset.
+ */
+export interface OATupleFields {
+	house_number: string
+	street: string
+	locality: string
+	postcode: string
+}
+
+export interface ReadOATuplesOptions<T> {
+	/**
+	 * Stop after this many distinct tuples — the `break` closes the reader and releases the archive, which is what the
+	 * GB-scale countrywide extracts need.
+	 */
+	limit?: number
+	/**
+	 * Drop rows without a postcode (reversed-order and balance readers, where the postcode drives the rendering).
+	 */
+	requirePostcode?: boolean
+	/**
+	 * Fold the postcode into the dedup key (`fr-order`'s key). The default key is
+	 * `${house_number}|${street}|${locality}`, lower-cased.
+	 */
+	dedupIncludesPostcode?: boolean
+	/**
+	 * Shape the recipe's tuple from the base fields, the raw record, and the (lower-cased) dedup key.
+	 */
+	extra: (fields: OATupleFields, row: CSVRecord, key: string) => T
+}
+
+/**
+ * Stream distinct tuples out of a cached OA zip — the reader the OA-skeleton recipes (`street-affix`, `unit`,
+ * `country-balanced`, `fr-order`) share. Field reads, filters, dedup keys and row order are exactly what each recipe's
+ * local copy did, so a recipe's slice stays byte-identical.
+ *
+ * @category CSV
+ */
+export async function readOATuples<T>(source: OATupleSource, options: ReadOATuplesOptions<T>): Promise<T[]> {
+	const tuples: T[] = []
+	const seen = new Set<string>()
+
+	for await (const row of readZippedCSVRecords(source.zip, source.csv)) {
+		if (options.limit !== undefined && tuples.length >= options.limit) break
+
+		const street = row.street ?? ""
+		const locality = row.city ?? ""
+		const house_number = row.number ?? ""
+		const postcode = row.postcode ?? ""
+
+		if (!street || !locality || !house_number) continue
+
+		if (options.requirePostcode && !postcode) continue
+
+		const key = (
+			options.dedupIncludesPostcode
+				? `${house_number}|${street}|${locality}|${postcode}`
+				: `${house_number}|${street}|${locality}`
+		).toLowerCase()
+
+		if (seen.has(key)) continue
+		seen.add(key)
+
+		tuples.push(options.extra({ house_number, street, locality, postcode }, row, key))
+	}
+
+	return tuples
+}
+
+/**
+ * Stream-parse a tuples JSONL file, yielding each parsed object (blank/invalid lines skipped).
+ */
+export function readTuples(input: PathBuilderLike): AsyncSequence<SliceTuple> {
+	// TextSpliterator (not JSONSpliterator) keeps the reader's established tolerance: malformed lines are skipped rather
+	// than rejecting the sequence. These operators fuse into the source's pull loop instead of adding an async-generator
+	// frame per tuple.
+	return TextSpliterator.fromAsync(input)
+		.map((line) => line.trim())
+		.filter((line) => Boolean(line))
+		.map((line) => tryParsingJSON<SliceTuple>(line))
+		.filter((tuple) => tuple !== null)
+		.map((tuple) => tuple!)
+}
+
+/**
+ * A canonical row as the recipes assemble it, before `alignRow` turns it into a `LabeledRow`.
+ */
+export interface CanonicalSliceRow {
+	raw: string
+	components: Record<string, string>
+	country: string
+	locale?: string
+	source: string
+	source_id: string
+	corpus_version?: string
+	license?: string
+}
+
+/**
+ * Run a canonical row through `alignRow` and, on success, write the `LabeledRow` (+ `synth_method` / `synth_base_id`)
+ * as one JSONL line. Returns true if emitted, false if alignment quarantined it.
+ */
+export function alignAndWrite(
+	write: (line: string) => void,
+	canonical: CanonicalSliceRow,
+	synthMethod: string,
+	synthBaseID: string | null = null
+): boolean {
+	const aligned = alignRow(canonical as Parameters<typeof alignRow>[0])
+
+	if (!aligned.row) return false
+	write(JSON.stringify({ ...aligned.row, synth_method: synthMethod, synth_base_id: synthBaseID }) + "\n")
+
+	return true
+}
+
+/**
+ * Parsed options a recipe's `run` receives. Common fields + the union of recipe-specific flags.
+ */
+export interface SliceRecipeOpts {
+	output: string
+	seed: number
+	variants: number
+	input?: string
+	count?: number
+	golden?: boolean
+	sourceName?: string
+	// recipe-specific (each recipe reads only what it needs):
+	houseNumberProb?: number
+	pmbRatio?: number
+	militaryRatio?: number
+	reversedFraction?: number
+	edgesDir?: string
+	country?: string
+	intlFraction?: number
+	/**
+	 * `german`: fraction of NATIVE-order rows rendered with no commas at all (`Neusser Str. 12 Nippes 50733 Köln`), the
+	 * single-line register #1946 found the model loses the district on. Default 0.3.
+	 */
+	commaFreeFraction?: number
+	/**
+	 * `german`: fraction of rows that carry a WOF Ortsteil of the tuple's locality as `dependent_locality`. Default 0.3;
+	 * 0 when no admin database is readable.
+	 */
+	ortsteilFraction?: number
+	/**
+	 * `german`: the WOF admin database the Ortsteil pool is read from. Default
+	 * `$MAILWOMAN_DATA_ROOT/wof/admin-global-priority-importance.db`.
+	 */
+	adminDB?: string
+	/**
+	 * `locale`: fraction of rows that append an explicit country surface form + a `country` component. Default 0.
+	 */
+	countryFraction?: number
+	/**
+	 * `locale`: tri-state override of the per-part `districtAsLocality` mapping for this invocation. `undefined` (flag
+	 * absent) leaves each `COUNTRY_SOURCES` part's own value untouched — every existing locale build stays
+	 * byte-identical. `true`/`false` forces that value on every part read this run. ES's pedanía slice
+	 * (`synth-es-pedania`) additionally uses `true` to select {@link LocaleCountrySource.pedaniaParts} instead of the
+	 * default `parts` — see `locale.ts`.
+	 */
+	districtAsLocality?: boolean
+	bareProb?: number
+	hnProb?: number
+	communes?: string
+	/**
+	 * `fr-lieudit`: BAN `adresses-<dept>.csv` directory. Default `$MAILWOMAN_DATA_ROOT/corpus/sources/ban`.
+	 */
+	banDir?: string
+	multilocaleCount?: number
+	/**
+	 * `fr-fragment` / `no-fragment` / `no-street-led`: the eval board's reserved street-surface list. REQUIRED for those
+	 * recipes — a slice that trains on its own eval set measures memorization. See their docstrings.
+	 */
+	excludeSurfaces?: string
+	/**
+	 * `no-fragment`: share of rows that are counter-distribution (bare locality OR bare postcode).
+	 */
+	counterProb?: number
+	/**
+	 * `no-fragment` knob 3: emit N copies of each street+number row whose number has >= longNumberMinDigits digits
+	 * (oversample the failing long-number class). Default 1 = no boost.
+	 */
+	longNumberBoost?: number
+	/**
+	 * `no-fragment` knob 3: minimum digit count for a number to count as "long" and be boosted. Default 3.
+	 */
+	longNumberMinDigits?: number
+	/**
+	 * `sub-venue`: the sub-venue lexicon JSON. Default = the committed `corpus/data/sub-venue-lexicon.json`, resolved
+	 * through the package manifest so it works from the source tree and from `out/`.
+	 */
+	lexicon?: string
+	/**
+	 * `sub-venue`: directory of `sub-venue-extract` JSONLs, one per region. Default
+	 * `$MAILWOMAN_DATA_ROOT/sub-venue/extracts`.
+	 */
+	extractsDir?: string
+	/**
+	 * `sub-venue`: the `poi.db` spatial layer, read for the en-US and fr-FR venue + confound pools (the two of poi.db's
+	 * four countries this slice has legs for). Default `$MAILWOMAN_DATA_ROOT/poi/poi.db`.
+	 */
+	poiDB?: string
+	/**
+	 * `sub-venue`: GB/US/FR address-context tuples JSONL. Default the house-venue v3 tuples
+	 * (`$MAILWOMAN_DATA_ROOT/corpus/intermediate/house-venue-tuples-v3.jsonl`); DE and ES read OpenAddresses directly.
+	 */
+	subVenueTuples?: string
+	/**
+	 * `sub-venue`: share of emitted rows that are confound NEGATIVES. Default 0.3.
+	 */
+	negativeFraction?: number
+}
+
+/**
+ * Tally a recipe returns.
+ */
+export interface SliceStats {
+	read?: number
+	emitted: number
+	skipped: number
+	/**
+	 * Rows dropped because their street SURFACE is reserved by an eval board (`--exclude-surfaces`). Separate from
+	 * `skipped` on purpose: a nonzero value is the audit trail that the train/eval split actually fired. Zero when a
+	 * recipe has no board split.
+	 */
+	contaminated?: number
+}
+
+/**
+ * A single declared recipe-specific option flag (for the command's --help).
+ */
+export interface SliceRecipeOption {
+	flag: string
+	description: string
+}
+
+/**
+ * A slice recipe: its identity, input mode, and its synthesis `run`.
+ */
+export interface CorpusRecipe {
+	/**
+	 * Recipe id, e.g. "street", "po-box" — the `<recipe>` positional.
+	 */
+	name: string
+	/**
+	 * One-line description for `--list` / help.
+	 */
+	description: string
+	/**
+	 * `tuples` reads `--input` JSONL; `generate` self-generates `--count` rows.
+	 */
+	mode: "tuples" | "generate"
+	/**
+	 * Recipe-specific flags this recipe honors (documentation only).
+	 */
+	options?: SliceRecipeOption[]
+	/**
+	 * Do the build: create the recipe's PRNG from `opts.seed` (its LEGACY generator — `makeLcg` or `makeMulberry32` — for
+	 * byte-reproducibility), synthesize, and emit each row via `write`.
+	 */
+	run(opts: SliceRecipeOpts, write: (line: string) => void): Promise<SliceStats>
+}

package/lib/recipes/si-bare-village.ts

@@ -0,0 +1,97 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `si-bare-village` — the Slovenian no-street counter-slice (#901 run-2). Slovenia's rural
+ *   addressing has NO street line: the village name is the street-level token and repeats as the
+ *   locality ("Zabiče 8, 6250 Zabiče"). The v1.9.8 check FALSIFIED the fr-bare-street slice alone on
+ *   exactly this class (SI resolve −3.4pp; "Apače 108" split into street "Apače 10" + house "8") —
+ *   the bare-street boundary lesson generalizes onto a form where the leading name must keep its
+ *   number whole and the trailing mention must stay locality-bound. This slice is the paired
+ *   counter-distribution: same lesson ("name before number, comma, then admin"), opposite polarity
+ *   on the trailing mention.
+ *
+ *   Three real-order templates cycle per tuple (mirrors the coord-golden orders so the eval and the
+ *   training distribution agree):
+ *
+ *   1. canonical  "«V» «n», «pc» «V»"
+ *   2. bare       "«V» «n», «V»"        (no postcode — the anchor-free form, the fr-bare lesson)
+ *   3. pc-first   "«pc» «V», «V» «n»"
+ *
+ *   Gold spans: leading «V» = street (matches OA ground truth — the village IS the address line),
+ *   «n» = house_number (NEVER split mid-digits), «pc» = postcode (never swallowing the neighbor),
+ *   trailing «V» = locality (the binding the resolver needs).
+ */
+
+import { alignAndWrite, makeMulberry32, readTuples, type CorpusRecipe, sliceSourceID } from "#recipes/scaffold"
+
+/**
+ * Slice 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 const siBareVillageRecipe: CorpusRecipe = {
+	name: "si-bare-village",
+	description: "SI no-street village form (#901 run-2): '«V» «n», «pc» «V»' — the fr-bare-street counter-distribution",
+	mode: "tuples",
+	async run(opts, write) {
+		makeMulberry32(opts.seed)
+		let read = 0
+		let emitted = 0
+		let skipped = 0
+
+		for await (const t of readTuples(opts.input!)) {
+			read++
+			const village = String(t.locality ?? "").trim()
+			const number = String(t.number ?? "").trim()
+			const postcode = String(t.postcode ?? "").trim()
+
+			if (!village || !number || !postcode) {
+				skipped++
+
+				continue
+			}
+
+			const order = read % 3
+			let raw: string
+
+			const components: Record<string, string> = {
+				street: village,
+				house_number: number,
+				locality: village,
+			}
+
+			if (order === 0) {
+				components.postcode = postcode
+				raw = `${village} ${number}, ${postcode} ${village}`
+			} else if (order === 1) {
+				raw = `${village} ${number}, ${village}`
+			} else {
+				components.postcode = postcode
+				raw = `${postcode} ${village}, ${village} ${number}`
+			}
+
+			const source_id = sliceSourceID("synth-si-bare-village", { ...components, o: String(order), v: String(read) })
+
+			const canonical = {
+				raw,
+				components,
+				country: "SI",
+				locale: "sl-SI",
+				source: "synth-si-bare-village",
+				source_id,
+				corpus_version: "0.9.9",
+				license:
+					"Synthetic — si-bare-village; (village, number, postcode) from OpenAddresses SI (per-source attribution in the model card)",
+			}
+
+			if (alignAndWrite(write, canonical, "si-bare-village")) {
+				emitted++
+			} else {
+				skipped++
+			}
+		}
+
+		return { read, emitted, skipped }
+	},
+}