/**
 * @copyright Sister Software
 * @license AGPL-3.0
 * @author Teffen Ellis, et al.
 *
 *   Train the #244 coarse-placer: a multinomial logistic-regression over the hashed char-n-gram +
 *   script features ({@link featurize}), via plain SGD. CPU-only, a few minutes — no GPU/Modal.
 *   After training, fits a single temperature on val (NLL minimization) for calibrated confidence.
 *   Writes a `meta.json` + `weights.bin` (Float32, row-major [class][feature]) artifact.
 *
 *   Run: `mailwoman placer train [--epochs 12] [--lr 0.1] [--l2 1e-6] [--out
 *   $MAILWOMAN_DATA_ROOT/coarse-placer/model]`
 */
/**
 * Options for {@linkcode trainCoarsePlacer}.
 */
export interface TrainCoarsePlacerOptions {
    /**
     * SGD epochs. Default 12.
     */
    epochs?: number;
    /**
     * Initial learning rate (decays per epoch). Default 0.1.
     */
    lr?: number;
    /**
     * L2 regularization. Default 1e-6.
     */
    l2?: number;
    /**
     * Artifact output dir. Default `$MAILWOMAN_DATA_ROOT/coarse-placer/model`.
     */
    out?: string;
    /**
     * Dataset dir (`{train,val}.jsonl`). Default `<repo>/data/coarse-placer`.
     */
    data?: string;
}
/**
 * Result of {@linkcode trainCoarsePlacer}.
 */
export interface TrainCoarsePlacerResult {
    outDir: string;
    trainRows: number;
    valRows: number;
    temperature: number;
    valNLL: number;
}
/**
 * Coarse-placer SGD trainer — see the module doc.
 */
export declare function trainCoarsePlacer(options?: TrainCoarsePlacerOptions, report?: (line: string) => void): Promise<TrainCoarsePlacerResult>;
//# sourceMappingURL=train.d.ts.map