84 lines
3.9 KiB
TypeScript
84 lines
3.9 KiB
TypeScript
import { describe, expect, test } from "bun:test";
|
|
import { renderReport, resultsJson, type BenchRun } from "../../src/bench/report.ts";
|
|
import type { SongScore } from "../../src/eval/score.ts";
|
|
|
|
const mr = (f1: number | null, oct = 0) => ({ tp: 0, fp: 0, fn: 0, precision: f1, recall: f1, f1, octaveErrors: oct, matches: [] as [number, number][] });
|
|
function score(gf1: number, vocals: number | null, flags = { oracleStems: false, oracleTempo: false, oracleNotes: false }): SongScore {
|
|
return {
|
|
flags,
|
|
guitarMerged: { ...mr(gf1, 1), pass: gf1 >= 0.85 },
|
|
partSplit: null,
|
|
perRole: { guitar: mr(gf1), vocals: mr(vocals), bass: mr(0.5) },
|
|
drumsPerClass: null,
|
|
rhythm: { factor: 1, beatOffset: 0, onBeat: 0.9, durationOk: null },
|
|
stringAgreement: null,
|
|
spuriousRate: 0.1,
|
|
gpPitch: { ...mr(gf1), tp: 0, fp: 0, fn: 1 },
|
|
barAgreement: { refBars: 8, outBars: 8, timeSigMatch: true },
|
|
};
|
|
}
|
|
const modes = [
|
|
{ name: "full", flags: { oracleStems: false, oracleTempo: false, oracleNotes: false } },
|
|
{ name: "oracle-notes", flags: { oracleStems: true, oracleTempo: true, oracleNotes: true } },
|
|
];
|
|
const mk = (g: number): BenchRun => ({
|
|
runId: "fx", date: "2026-10-30", modes,
|
|
results: [
|
|
{ tier: "t1a", seed: 1, mode: "full", score: score(g, null), seconds: 1 },
|
|
{ tier: "t1a", seed: 2, mode: "full", score: score(g, 0.4), seconds: 2 },
|
|
{ tier: "t1a", seed: 1, mode: "oracle-notes", score: score(1, 1, modes[1]!.flags), seconds: 1 },
|
|
],
|
|
});
|
|
|
|
describe("renderReport", () => {
|
|
test("column order, n/a, caveat", () => {
|
|
const md = renderReport(mk(0.9));
|
|
const header = md.split("\n").find((l) => l.startsWith("| Mode"))!;
|
|
expect(header.indexOf("Guitar merged F1")).toBeLessThan(header.indexOf("Vocals F1"));
|
|
expect(header.indexOf("Vocals F1")).toBeLessThan(header.indexOf("Bass F1"));
|
|
expect(md).toContain("n/a");
|
|
expect(md).toContain("Voice Oohs");
|
|
expect(md).toContain("fx");
|
|
});
|
|
test("tier pass line", () => {
|
|
expect(renderReport(mk(0.9))).toContain("PASS");
|
|
const bad = renderReport(mk(0.5));
|
|
expect(bad).toContain("FAIL");
|
|
expect(bad).not.toContain("Tier t1a: PASS");
|
|
});
|
|
test("per-mode tick uses the median rule with a k/n pass count", () => {
|
|
const gs = [0.9, 0.9, 0.9, 0.9, 0.3];
|
|
const run: BenchRun = { runId: "fx", date: "d", modes: [modes[0]!], results: gs.map((g, i) => ({ tier: "t1a" as const, seed: i, mode: "full", score: score(g, null), seconds: 1 })) };
|
|
const row = renderReport(run).split("\n").find((l) => l.startsWith("| full"))!;
|
|
expect(row).toContain("✓ 4/5");
|
|
expect(row).not.toContain("✗");
|
|
const bad = renderReport({ ...run, results: gs.map((g, i) => ({ tier: "t1a" as const, seed: i, mode: "full", score: score(g === 0.9 ? 0.3 : 0.9, null), seconds: 1 })) });
|
|
expect(bad.split("\n").find((l) => l.startsWith("| full"))).toContain("✗ 1/5");
|
|
});
|
|
test("secondary metric columns: GP-file F1, bar agreement, techniques n/a", () => {
|
|
const md = renderReport(mk(0.9));
|
|
const header = md.split("\n").find((l) => l.startsWith("| Mode"))!;
|
|
expect(header).toContain("GP-file guitar F1");
|
|
expect(header).toContain("Bar count match");
|
|
expect(header).toContain("Techniques");
|
|
expect(md).toContain("n/a (tier 1 has no technique marks)");
|
|
});
|
|
test("role with empty truth renders n/a with spurious count", () => {
|
|
const run = mk(0.9);
|
|
const s = run.results[0]!.score;
|
|
s.perRole.bass = { ...mr(0), tp: 0, fp: 3, fn: 0 };
|
|
const md = renderReport({ ...run, results: [run.results[0]!] });
|
|
expect(md).toContain("n/a (3 spurious)");
|
|
expect(md).not.toContain("0.00 (0.00");
|
|
});
|
|
});
|
|
|
|
test("results JSON omits per-note matches", () => {
|
|
const run = mk(0.9);
|
|
run.results[0]!.score.guitarMerged.matches = [[0, 1]];
|
|
const json = resultsJson(run);
|
|
expect(json).not.toContain('"matches"');
|
|
expect(JSON.parse(json).results).toHaveLength(3);
|
|
expect(JSON.stringify(run)).toContain('"matches"'); // in-memory run (and score.json) keep them
|
|
});
|