/**
 * bench/harness.ts — run a benchmark and A/B two models.
 *
 * The harness drives a model over held-out token sequences, scores next-token
 * prediction with {@link metrics}, and produces a {@link BenchmarkReport}. It
 * also offers a one-call {@link trainAndBenchmark} that builds a fresh
 * `EvermindLM` from a corpus, holds out a slice, trains on the rest, and scores
 * — the path the Studio uses to benchmark a model in the browser.
 */
import type { AsyncLogitsModel, BenchmarkOptions, BenchmarkReport, ComparisonReport, LogitsModel, TrainAndBenchmarkOptions, TrainAndBenchmarkResult } from "./types.js";
/**
 * Benchmark a model (synchronous forward) over held-out token sequences.
 * Sequences shorter than 2 tokens have no prediction target and are skipped.
 */
export declare function benchmarkModel(model: LogitsModel, sequences: number[][], opts?: BenchmarkOptions): BenchmarkReport;
/** Benchmark a model whose forward pass is asynchronous (e.g. a WebGPU backend). */
export declare function benchmarkModelAsync(model: AsyncLogitsModel, sequences: number[][], opts?: BenchmarkOptions): Promise<BenchmarkReport>;
/** A/B two models on the same eval set. Lower perplexity wins (the primary metric). */
export declare function compareModels(candidate: LogitsModel, baseline: LogitsModel, sequences: number[][], opts?: BenchmarkOptions): ComparisonReport;
/** Build a {@link ComparisonReport} from two already-computed scorecards. */
export declare function compareReports(candidate: BenchmarkReport, baseline: BenchmarkReport): ComparisonReport;
/** Split a corpus into next-token training sequences (one per sentence). */
export declare function corpusToSequences(corpus: string, codec: {
    encode(text: string): number[];
}): number[][];
/** Benchmark an existing model against a raw-text corpus (tokenized per sentence). */
export declare function benchmarkText(model: LogitsModel, codec: {
    encode(text: string): number[];
}, corpus: string, opts?: BenchmarkOptions): BenchmarkReport;
/**
 * Train a fresh `EvermindLM` on a corpus and benchmark it on a held-out slice —
 * the canonical "build + score a model from text" path. A real benchmark must
 * score data the model never trained on, so `heldOutRatio` is clamped to leave
 * at least one eval sequence and one train sequence.
 */
export declare function trainAndBenchmark(corpus: string, opts?: TrainAndBenchmarkOptions): TrainAndBenchmarkResult;
//# sourceMappingURL=harness.d.ts.map