/** * @copyright Sister Software * @license AGPL-3.0 * @author Teffen Ellis, et al. * * Accuracy cost of the #244 coarse-placer int8 quantization (milestone 3). Runs the held-out test * split through the fp32 model and the int8 model (dequantized inline) and reports overall + per- * class accuracy for each, the delta, prediction-agreement rate, and confidence MAE. Gate: int8 * within ~1pp of fp32 (the milestone target). * * Run: `mailwoman placer eval quant-compare [--fp32 ] [--int8 ] [--abstain 0.5]` */ /** * Options for {@linkcode evalQuantCompare}. */ export interface EvalQuantCompareOptions { /** * Fp32 artifact dir. Default `$MAILWOMAN_DATA_ROOT/coarse-placer/model`. */ fp32?: string; /** * Int8 artifact dir. Default `$MAILWOMAN_DATA_ROOT/coarse-placer/model-int8`. */ int8?: string; /** * Abstention threshold. Default 0.5. */ abstain?: number; /** * Dataset dir (`test.jsonl`). Default `/data/coarse-placer`. */ data?: string; } /** * Result of {@linkcode evalQuantCompare}. */ export interface EvalQuantCompareResult { n: number; /** * Fp32 overall accuracy in percent. */ accFp32: number; /** * Int8 overall accuracy in percent. */ accInt8: number; /** * Whether int8 is within 1pp of fp32 (the gate). */ pass: boolean; } /** * Coarse-placer int8-vs-fp32 comparison — see the module doc. Emits the report to stdout. */ export declare function evalQuantCompare(options?: EvalQuantCompareOptions): Promise; //# sourceMappingURL=eval-quant-compare.d.ts.map