@mailwoman/corpus 9.4.0 → 10.1.0

package/lib/tools/index.ts

@@ -1,28 +0,0 @@
-/**
- * @copyright Sister Software
- * @license AGPL-3.0
- * @author Teffen Ellis, et al.
- *
- *   Corpus operator tools — the `run()`-style modules behind `mailwoman corpus …` commands. No argv,
- *   no `process.exit`: commands own parsing, rendering, and exit codes.
- */
-
-export * from "#tools/align-slice"
-export * from "#tools/audit"
-export * from "#tools/corpus-stats"
-export * from "#tools/ingest-csv"
-export * from "#tools/overlay-manifest"
-export * from "#tools/fetch/download"
-export * from "#tools/fetch/index"
-export * from "#tools/golden-expand"
-export * from "#tools/golden-promote"
-export * from "#tools/golden-relabel-street"
-export * from "#tools/jsonl-to-parquet"
-export * from "#tools/lint-slice"
-export * from "#tools/lint-slice-vocab"
-export * from "#tools/slice-kryptonite"
-export * from "#tools/overture-subvenue"
-export * from "#tools/postcode-triples"
-export * from "#tools/sub-venue-lexicon"
-export * from "#tools/sub-venue-promotions"
-export * from "#tools/slice-translit"

package/lib/tools/lint-slice-vocab.ts

@@ -1,388 +0,0 @@
-/**
- * @copyright Sister Software
- * @license AGPL-3.0
- * @author Teffen Ellis, et al.
- *
- *   #511 base-consistency lint, GENERALIZED + COUNTRY-SCOPED (v2) — any synthetic slice vs the base.
- *
- *   Ported from scripts/lint-slice-vocab.py (pyarrow → @duckdb/node-api); behavior preserved
- *   byte-for-byte (same flags, same stdout, same verdicts). The base-root default routes through
- *   `dataRootPath` so the lab `/mnt/playpen` literal stays in its one home
- *   (core/utils/data-root.ts) and `$MAILWOMAN_DATA_ROOT` is honored; with the env unset it equals
- *   the Python default.
- *
- *   The #511 lesson: a synthetic slice must not label a token a tag the BASE dominantly labels
- *   something else, or training gets conflicting gradients on the same token and the minority (the
- *   slice) loses. This reads a slice's own (token -> tag) and checks each token against the base.
- *
- *   WHY v2 IS COUNTRY-SCOPED + FULL-COUNT (the night-2026-06-18 lesson, learned the hard way over
- *   three tries): a token's correct tag is COUNTRY-specific — "Paris" is locality in FR data and
- *   street in US "Paris Ave"; "Marion" is a US town AND many US "Marion" streets. So:
- *
- *   1. A cross-COUNTRY aggregate mis-judges any country-specific token (v1 uniform AND a proportional
- *        retry both false-flagged FR cities as "street" from US street-contexts).
- *   2. A SMALL sample is street-BIASED regardless, because the street sources (tiger 39 + nad 378 parts)
- *        dwarf the locality sources (a small US-scoped spot-check read Indianapolis 54% street vs
- *        its true 219700:29 LOCALITY). The fix: tally each slice token's base tag SCOPED to the
- *        country the slice uses it in (the base has a `country` column), over a LARGE/FULL scan
- *        (`fraction`, default 1.0). Pure-numeric tokens excluded (house_number/postcode are
- *        context-determined). An affix-split flag (slice street_suffix/_prefix vs base "street") is
- *        EXPECTED — the loader's affix-relabel handles it; weigh those separately.
- *
- *   Usage: mailwoman dev lint slice-vocab --slice <slice.parquet>
- *   [--base-version v0.5.0] [--base-root <dir>] [--fraction 1.0] [--threshold 0.7] [--min-count
- *   50]
- */
-
-import { dataRootPath } from "@mailwoman/core/data-root"
-import { readDirectory } from "@mailwoman/core/fs/readers"
-import { tryParsingJSON } from "@mailwoman/core/json"
-import { pyRound } from "@mailwoman/core/numeric"
-import { join } from "path-ts"
-
-import { connectDuckDB, type DuckDBConnection } from "#utils/parquet"
-
-/**
- * A column-projected base/slice row: parallel token + label lists plus the row's country.
- */
-interface CorpusRow {
-	tokens: string[]
-	labels: string[]
-	country: string | null
-}
-
-/**
- * Strip a BIO prefix ("B-"/"I-") off a label, matching the Python `strip_bio`.
- */
-function stripBIO(label: string): string {
-	const head = label.slice(0, 2)
-
-	return head === "B-" || head === "I-" ? label.slice(2) : label
-}
-
-/**
- * Python `str.isdigit()`: non-empty and every character a Unicode digit. Pure-numeric tokens (house_number / postcode)
- * are context-determined, not lexical vocab, so they're excluded. `\p{Nd}` covers the decimal digits these address
- * corpora actually contain.
- */
-function isDigit(token: string): boolean {
-	return token.length > 0 && /^\p{Nd}+$/u.test(token)
-}
-
-/**
- * Format a fraction as a whole-percent string the way Python's `:.0%` does, e.g. 0.73 -> "73%".
- */
-function pct(frac: number): string {
-	return `${pyRound(frac * 100)}%`
-}
-
-/**
- * Format a float the way a Python f-string renders it: integer-valued floats keep one decimal (1.0 -> "1.0"),
- * everything else is its shortest decimal (0.5 -> "0.5"). Used for the `fraction` echo so the banner matches the Python
- * print.
- */
-function formatPyFloat(n: number): string {
-	return Number.isInteger(n) ? n.toFixed(1) : String(n)
-}
-
-/**
- * Left-justify to a minimum width with spaces, matching Python's `{value:N}` string field.
- */
-function pad(value: string, width: number): string {
-	return value.padEnd(width)
-}
-
-/**
- * The dominant tag of a counter: (tag, total, fraction). Empty counter -> ("", 0, 0.0). Ties go to the first-inserted
- * tag, mirroring `Counter.most_common(1)` (stable on equal counts).
- */
-function dominant(counter: Map<string, number>): [string, number, number] {
-	let total = 0
-	let bestTag = ""
-	let bestCount = -1
-
-	for (const [tag, count] of counter) {
-		total += count
-
-		if (count > bestCount) {
-			bestCount = count
-			bestTag = tag
-		}
-	}
-
-	if (total === 0) return ["", 0, 0]
-
-	return [bestTag, total, bestCount / total]
-}
-
-/**
- * Bump a (key -> count) tally, creating the inner counter on first sight.
- */
-function bump(table: Map<string, Map<string, number>>, key: string, sub: string): void {
-	let counter = table.get(key)
-
-	if (!counter) {
-		counter = new Map()
-		table.set(key, counter)
-	}
-
-	counter.set(sub, (counter.get(sub) ?? 0) + 1)
-}
-
-/**
- * Read a corpus parquet into rows, projecting only tokens/labels/country. The list columns ride out as JSON text
- * (DuckDB `to_json`) — the same trick the gazetteer builders use for nested columns — and parse back to string arrays
- * here.
- */
-async function readRows(con: DuckDBConnection, path: string): Promise<CorpusRow[]> {
-	const result = await con.runAndReadAll(
-		`SELECT to_json(tokens) AS tokens, to_json(labels) AS labels, country FROM read_parquet('${path}')`
-	)
-
-	const raw = result.getRowObjects() as Array<{ tokens: unknown; labels: unknown; country: unknown }>
-	const rows: CorpusRow[] = []
-
-	for (const r of raw) {
-		const tokens = tryParsingJSON(String(r.tokens))
-		const labels = tryParsingJSON(String(r.labels))
-
-		if (!Array.isArray(tokens) || !Array.isArray(labels)) continue
-
-		rows.push({
-			tokens: tokens as string[],
-			labels: labels as string[],
-			country: r.country == null ? null : String(r.country),
-		})
-	}
-
-	return rows
-}
-
-/**
- * Read just the first row's `source` value — used to group base parts for a proportional slice.
- */
-async function readSource(con: DuckDBConnection, path: string): Promise<string> {
-	const result = await con.runAndReadAll(`SELECT source FROM read_parquet('${path}') LIMIT 1`)
-	const rows = result.getRowObjects() as Array<{ source: unknown }>
-
-	return rows.length ? String(rows[0]!.source) : ""
-}
-
-/**
- * Non-recursive `*.parquet` glob, sorted lexicographically — the Python `sorted(glob.glob(...))`.
- */
-async function globParquet(dir: string): Promise<string[]> {
-	let names: string[]
-
-	try {
-		names = await readDirectory(dir)
-	} catch {
-		return []
-	}
-
-	return names
-		.filter((f) => f.endsWith(".parquet"))
-		.map((f) => join(dir, f))
-		.toSorted()
-}
-
-/**
- * Options for {@linkcode lintSliceVocab}.
- */
-export interface LintSliceVocabOptions {
-	/**
-	 * The slice parquet to lint.
-	 */
-	slice: string
-	/**
-	 * Base corpus version. Default `v0.5.0`.
-	 */
-	baseVersion?: string
-	/**
-	 * Base corpus root. Default `$MAILWOMAN_DATA_ROOT/corpus/versioned`.
-	 */
-	baseRoot?: string
-	/**
-	 * Base-majority confidence floor for a contradiction. Default 0.7.
-	 */
-	threshold?: number
-	/**
-	 * Minimum base support to judge a token. Default 50.
-	 */
-	minCount?: number
-	/**
-	 * Fraction of base parts to scan (proportional per-source slice below 1.0). Default 1.0.
-	 */
-	fraction?: number
-}
-
-/**
- * One contradiction row: token, slice tag, base tag, base fraction, base total.
- */
-export type SliceVocabRow = [token: string, sliceTag: string, baseTag: string, baseFrac: number, baseTotal: number]
-
-/**
- * Findings summary returned by {@linkcode lintSliceVocab}.
- */
-export interface LintSliceVocabSummary {
-	/**
-	 * Real contradictions — the command exits 1 when nonzero.
-	 */
-	errors: number
-	/**
-	 * Affix-split rows (EXPECTED — the loader's affix-relabel handles them).
-	 */
-	warnings: number
-	findings: { contradictions: SliceVocabRow[]; affixSplits: SliceVocabRow[] }
-}
-
-/**
- * Lint a synthetic slice's (token → tag) vocabulary against the base corpus, country-scoped.
- */
-export async function lintSliceVocab(options: LintSliceVocabOptions): Promise<LintSliceVocabSummary> {
-	const baseVersion = options.baseVersion ?? "v0.5.0"
-	const baseRoot = options.baseRoot ?? dataRootPath("corpus", "versioned")
-	const threshold = options.threshold ?? 0.7
-	const minCount = options.minCount ?? 50
-	const fraction = options.fraction ?? 1
-
-	const con = await connectDuckDB()
-
-	// 1. the slice's own (token -> dominant tag) + the COUNTRIES it uses each token in
-	const sliceRows = await readRows(con, options.slice)
-	const sliceTags = new Map<string, Map<string, number>>()
-	const sliceCountries = new Map<string, Set<string | null>>()
-
-	for (const { tokens, labels, country } of sliceRows) {
-		const n = Math.min(tokens.length, labels.length)
-
-		for (let i = 0; i < n; i++) {
-			const w = tokens[i]!
-			const l = labels[i]!
-
-			if (isDigit(w)) continue // numbers are context-determined (house_number/postcode), not lexical vocab
-			bump(sliceTags, w, stripBIO(l))
-			let set = sliceCountries.get(w)
-
-			if (!set) {
-				set = new Set()
-				sliceCountries.set(w, set)
-			}
-
-			set.add(country)
-		}
-	}
-
-	const sliceVocab = new Set(sliceTags.keys())
-
-	console.log(`slice: ${sliceRows.length} rows, ${sliceVocab.size} unique tokens`)
-
-	// 2. base parts — FULL by default; fraction<1 takes a proportional per-source slice (still big)
-	const trainDir = join(baseRoot, baseVersion, `corpus-${baseVersion}`, "train")
-	let parts = await globParquet(trainDir)
-
-	if (!parts.length) {
-		throw new Error("no base parts found")
-	}
-
-	if (fraction < 1) {
-		const bysrc = new Map<string, string[]>()
-
-		for (const p of parts) {
-			const src = await readSource(con, p)
-			let list = bysrc.get(src)
-
-			if (!list) {
-				list = []
-				bysrc.set(src, list)
-			}
-
-			list.push(p)
-		}
-
-		const sliced: string[] = []
-
-		for (const ps of bysrc.values()) {
-			const take = Math.max(2, pyRound(ps.length * fraction))
-
-			for (const p of ps.slice(0, take)) {
-				sliced.push(p)
-			}
-		}
-
-		parts = sliced
-	}
-
-	console.log(
-		`base ${baseVersion}: scanning ${parts.length} parts (fraction=${formatPyFloat(fraction)}), COUNTRY-scoped`
-	)
-
-	// 3. tally each slice token's base tag, SCOPED to the country the slice uses it in
-	const baseTags = new Map<string, Map<string, number>>()
-
-	for (let i = 0; i < parts.length; i++) {
-		const rows = await readRows(con, parts[i]!)
-
-		for (const { tokens, labels, country } of rows) {
-			const n = Math.min(tokens.length, labels.length)
-
-			for (let j = 0; j < n; j++) {
-				const w = tokens[j]!
-
-				if (sliceVocab.has(w) && sliceCountries.get(w)!.has(country)) {
-					bump(baseTags, w, stripBIO(labels[j]!))
-				}
-			}
-		}
-
-		if ((i + 1) % 100 === 0) {
-			console.log(`  ...${i + 1}/${parts.length} parts`)
-		}
-	}
-
-	// 4. compare; flag contradictions (affix-split is expected — surfaced but tagged)
-	const flagged: SliceVocabRow[] = []
-	const affix: SliceVocabRow[] = []
-
-	for (const w of sliceVocab) {
-		const [sTag] = dominant(sliceTags.get(w)!)
-		const [bTag, bTotal, bFrac] = dominant(baseTags.get(w) ?? new Map())
-
-		if (bTotal < minCount || !bTag || bTag === sTag || bFrac < threshold) continue
-		const row: SliceVocabRow = [w, sTag, bTag, bFrac, bTotal]
-
-		if ((sTag === "street_suffix" || sTag === "street_prefix") && bTag === "street") {
-			affix.push(row)
-		} else {
-			flagged.push(row)
-		}
-	}
-
-	const sections: Array<[string, SliceVocabRow[]]> = [
-		["CONTRADICTION", flagged],
-		["affix-split (EXPECTED — affix-relabel handles)", affix],
-	]
-
-	for (const [label, rows] of sections) {
-		if (!rows.length) continue
-		rows.sort((a, b) => b[4] - a[4] || b[3] - a[3])
-
-		console.log(`\n${label.startsWith("CONTRA") ? "⚠️ " : "· "}${rows.length} ${label}:`)
-
-		for (const [w, sTag, bTag, bFrac, bTotal] of rows) {
-			console.log(`  ${pad(w, 18)} slice=${pad(sTag, 14)} base=${bTag} (${pct(bFrac)}, n=${bTotal})`)
-		}
-	}
-
-	if (!flagged.length) {
-		console.log(
-			`\n✅ NO real contradictions (country-scoped, threshold ${pct(threshold)}, support ${minCount}) — slice base-consistent`
-		)
-	}
-
-	return {
-		errors: flagged.length,
-		warnings: affix.length,
-		findings: { contradictions: flagged, affixSplits: affix },
-	}
-}