/**
 * @copyright Sister Software
 * @license AGPL-3.0
 * @author Teffen Ellis, et al.
 *
 * Convert a jsonl of LabeledRow objects to a Parquet file matching the schema the corpus writes.
 *
 * DuckDB reproduces the PyArrow file's logical schema and column order. A PyArrow reader therefore sees an identical table.
 * The trainer reads Parquet by column name, so physical layout does not affect it.
 *
 * Schema: `PARQUET_COLUMNS` from `#parquet/schema`, the same list the native writer uses.
 *
 * Every row must contain the span triple. A row without it came from a producer that has not migrated.
 * The writer would silently drop the character-offset labels if it wrote that row. Fail loudly and report the row number.
 */
import type { PathBuilderLike } from "path-ts";
/**
 * Options for {@linkcode jsonlToParquet}.
 */
export interface JSONLToParquetOptions {
    /**
     * The labeled-row jsonl to convert.
     */
    input: PathBuilderLike;
    /**
     * The parquet file to write.
     */
    output: PathBuilderLike;
    /**
     * Parquet row-group size.
     *
     * Default 50000.
     */
    rowGroupSize?: number;
}
/**
 * Summary returned by {@linkcode jsonlToParquet}.
 */
export interface JSONLToParquetSummary {
    read: number;
    written: number;
    outPath: string;
}
/**
 * Convert a labeled-row jsonl to a v0.5.0-schema Parquet file.
 */
export declare function jsonlToParquet(options: JSONLToParquetOptions, report?: (line: string) => void): Promise<JSONLToParquetSummary>;
//# sourceMappingURL=jsonl-to-parquet.d.ts.map