/**
 * Skip-gram with negative sampling (word2vec SGNS) in pure TypeScript.
 *
 * Training is kept PURE — no subword features mixed into the gradient — so the
 * in-vocabulary word vectors stay semantic (deploy↔배포) rather than drifting
 * orthographic. Subword coverage for rare / unseen words is added afterwards as
 * a derived centroid layer (see buildSubwordCentroids), which leaves these
 * trained vectors untouched. Deterministic given `opts.seed`. Input is
 * pre-tokenized sentences (use the same tokenizer as memory search).
 */
import type { TrainedEmbeddings, Word2VecOptions } from "./types.js";
export declare const DEFAULT_W2V: Word2VecOptions;
export type TrainProgress = (epoch: number, epochs: number) => void;
export declare function trainWord2Vec(sentences: string[][], options?: Partial<Word2VecOptions>, onProgress?: TrainProgress): TrainedEmbeddings;