@exercode/problem-utils 2.3.0 → 2.3.1

package/dist/presets/evaluation.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"evaluation.cjs","names":["parseArgs","resolveCwds","checkProblemDirIsolation","path","judgeByStaticAnalysis","readProblemMarkdownFrontMatter","isSafeSubmissionOutputPath","fs","DecisionCode","parseCsvRecords","CsvParseError","parseCsv"],"sources":["../../src/presets/evaluation.ts"],"sourcesContent":["import fs from 'node:fs/promises';\nimport path from 'node:path';\n\nimport { checkProblemDirIsolation } from '../helpers/checkProblemDirIsolation.js';\nimport { judgeByStaticAnalysis } from '../helpers/judgeByStaticAnalysis.js';\nimport { parseArgs } from '../helpers/parseArgs.js';\nimport { CsvParseError, type CsvParseErrorReason, parseCsv, parseCsvRecords } from '../helpers/parseCsv.js';\nimport { printDebugBanner } from '../helpers/printDebugBanner.js';\nimport { printTestCaseResult } from '../helpers/printTestCaseResult.js';\nimport { isSafeSubmissionOutputPath } from '../helpers/readOutputFiles.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport {\n  printDebugCwdBanner,\n  printDebugExpectationFailureBanner,\n  resolveCwds,\n  type ResolvedCwd,\n} from '../helpers/resolveCwds.js';\nimport { DecisionCode } from '../types/decisionCode.js';\nimport type { TestCaseResult } from '../types/testCaseResult.js';\n\nconst DEFAULT_SUBMISSION_FILE_PATH = 'submission.csv';\nconst TEST_CASE_ID = 'evaluation';\nconst MAX_LISTED_IDS = 10;\n// Learner-controlled cells are echoed into the feedback, so clip them to keep the result line small.\nconst MAX_ECHOED_CELL_LENGTH = 50;\nconst CSV_PARSE_ERROR_MESSAGES: Record<CsvParseErrorReason, string> = {\n  unterminated_quote: '引用符（`\"`）が閉じられていません',\n  unexpected_quote: '引用符（`\"`）は値全体を囲む位置にしか書けません',\n  text_after_quoted_field: '閉じ引用符の後に区切り文字以外の文字があります',\n};\n\nexport interface EvaluationMetric {\n  /** Label shown with the score, e.g. `RMSLE`. */\n  label: string;\n  isHigherBetter: boolean;\n  /**\n   * Returns an error message when a predicted value cannot be scored (e.g. it is not a number).\n   * The message is appended to the row's line number and the offending value, so phrase it as a\n   * predicate such as `数値ではありません`. It also validates the ground truth when it is loaded.\n   */\n  validatePrediction?: (prediction: string) => string | undefined;\n  /** Computes the score from predictions and answers aligned by id. Both are raw CSV cell strings. */\n  compute: (predictions: readonly string[], answers: readonly string[]) => number;\n}\n\nexport interface EvaluationJudgePresetOptions {\n  /** CSV file holding the hidden ground truth, relative to the problem directory. */\n  answerFilePath: string;\n  /** CSV file the submission must contain. Defaults to `submission.csv`. */\n  submissionFilePath?: string;\n  /** Column that identifies each row in both CSV files. */\n  idColumn: string;\n  /** Column holding the answer in the ground truth and the prediction in the submission. */\n  targetColumn: string;\n  metric: EvaluationMetric;\n  /**\n   * Score required for acceptance: the submission is accepted when its score is at least this value\n   * for a higher-is-better metric and at most this value otherwise. Every valid submission is\n   * accepted when omitted.\n   */\n  acceptableScore?: number;\n}\n\nexport const evaluationMetrics = {\n  rmsle: {\n    label: 'RMSLE',\n    isHigherBetter: false,\n    validatePrediction: (prediction: string) =>\n      validateFiniteNumber(prediction) ?? (Number(prediction) < 0 ? '負の値です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(\n        mean(\n          predictions.map(\n            (prediction, index) => (Math.log1p(Number(prediction)) - Math.log1p(Number(answers[index]))) ** 2\n          )\n        )\n      ),\n  },\n  rmse: {\n    label: 'RMSE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(mean(predictions.map((prediction, index) => (Number(prediction) - Number(answers[index])) ** 2))),\n  },\n  mae: {\n    label: 'MAE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => Math.abs(Number(prediction) - Number(answers[index])))),\n  },\n  accuracy: {\n    label: 'Accuracy',\n    isHigherBetter: true,\n    validatePrediction: (prediction: string) => (prediction.trim() === '' ? '空です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => (prediction.trim() === answers[index]?.trim() ? 1 : 0))),\n  },\n} as const satisfies Record<string, EvaluationMetric>;\n\n/**\n * A preset for Kaggle-style model evaluation problems: the submission is a CSV of predictions for a\n * hidden test set, and the judge reports the metric value as `score` instead of comparing outputs.\n * The problem directory keeps the ground truth CSV, so it is never shipped to learners.\n *\n * @example\n * Create `judge.ts`:\n * ```ts\n * import { evaluationJudgePreset, evaluationMetrics } from '@exercode/problem-utils/presets/evaluation';\n *\n * await evaluationJudgePreset(import.meta.dirname, {\n *   answerFilePath: 'evaluation/answer.csv',\n *   idColumn: 'Id',\n *   targetColumn: 'TradePrice',\n *   metric: evaluationMetrics.rmsle,\n *   acceptableScore: 0.5,\n * });\n * ```\n *\n * Run `bun judge.ts model_answers/default` to judge one submission, or `bun judge.ts` to judge every\n * `<problemDir>/model_answers/*` and `<problemDir>/model_answers.fails/*` directory for debugging.\n */\nexport async function evaluationJudgePreset(problemDir: string, options: EvaluationJudgePresetOptions): Promise<void> {\n  const args = parseArgs(process.argv);\n  const { cwds, isDebugMode } = await resolveCwds(problemDir, args.cwd);\n\n  if (isDebugMode) {\n    const acceptedCwd = cwds.find((cwd) => cwd.expectedResult === 'accepted');\n    if (acceptedCwd) {\n      const isolationCheckResult = await checkProblemDirIsolation(problemDir, acceptedCwd, args.params);\n      if (!isolationCheckResult.passed) {\n        process.exitCode = 1;\n        return;\n      }\n    } else {\n      printDebugBanner([\n        '[DEBUG MODE] isolated problem directory check skipped',\n        '',\n        'No accepted model answer is available for checking that the copied judge still accepts a valid submission.',\n      ]);\n    }\n  }\n\n  const answer = await readAnswer(path.join(problemDir, options.answerFilePath), options);\n  for (const resolvedCwd of cwds) {\n    if (isDebugMode) printDebugCwdBanner(problemDir, resolvedCwd);\n    const result = await judgeSubmission(problemDir, resolvedCwd.cwd, answer, options);\n    printTestCaseResult({ testCaseId: TEST_CASE_ID, ...result });\n    if (isDebugMode && !matchesExpectedResult(resolvedCwd, result)) {\n      process.exitCode = 1;\n      printDebugExpectationFailureBanner(problemDir, resolvedCwd);\n    }\n  }\n}\n\nasync function judgeSubmission(\n  problemDir: string,\n  cwd: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Promise<Omit<TestCaseResult, 'testCaseId'>> {\n  const staticAnalysisResult = await judgeByStaticAnalysis(cwd, await readProblemMarkdownFrontMatter(problemDir));\n  if (staticAnalysisResult) return staticAnalysisResult;\n\n  const submissionFilePath = options.submissionFilePath ?? DEFAULT_SUBMISSION_FILE_PATH;\n  let submissionText: string;\n  try {\n    const resolvedPath = path.join(cwd, submissionFilePath);\n    // The harness reads the file as the trusted user, so refuse symlinks escaping the submission.\n    if (!(await isSafeSubmissionOutputPath(cwd, resolvedPath))) throw new Error('unsafe path');\n    submissionText = await fs.readFile(resolvedPath, 'utf8');\n  } catch {\n    return {\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      feedbackMarkdown: `\\`${submissionFilePath}\\`が見つかりません。予測結果のCSVファイルを\\`${submissionFilePath}\\`という名前で提出してください。`,\n    };\n  }\n\n  const predictions = readPredictions(submissionText, submissionFilePath, answer, options);\n  if (typeof predictions === 'string') {\n    return { decisionCode: DecisionCode.WRONG_ANSWER, feedbackMarkdown: predictions };\n  }\n\n  const answerValues = [...answer.values()];\n  const score = options.metric.compute(\n    [...answer.keys()].map((id) => predictions.get(id) as string),\n    answerValues\n  );\n  if (!Number.isFinite(score)) {\n    return {\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      feedbackMarkdown: `予測値が極端なため、${options.metric.label}を計算できませんでした。予測値を見直してください。`,\n    };\n  }\n  const isAccepted =\n    options.acceptableScore === undefined ||\n    (options.metric.isHigherBetter ? score >= options.acceptableScore : score <= options.acceptableScore);\n  return {\n    decisionCode: isAccepted ? DecisionCode.ACCEPTED : DecisionCode.WRONG_ANSWER,\n    score,\n    scoreLabel: options.metric.label,\n    feedbackMarkdown: buildScoreMarkdown(score, answerValues.length, isAccepted, options),\n  };\n}\n\nfunction buildScoreMarkdown(\n  score: number,\n  rowCount: number,\n  isAccepted: boolean,\n  options: EvaluationJudgePresetOptions\n): string {\n  const { label, isHigherBetter } = options.metric;\n  const precision = findDistinguishingPrecision(score, options.acceptableScore);\n  const acceptanceLine =\n    options.acceptableScore === undefined\n      ? ''\n      : `\\n合格基準: ${label} ${isHigherBetter ? '≧' : '≦'} ${formatScore(options.acceptableScore, precision)}（${isAccepted ? '達成' : '未達成'}）`;\n  return `| 指標 | スコア |\n| ---- | ------ |\n| ${label} | ${formatScore(score, precision)} |\n\n評価件数: ${rowCount.toLocaleString('en-US')}件${acceptanceLine}\n`;\n}\n\n/**\n * Finds the number of significant digits at which the score and the threshold render differently\n * (or stop changing), so the feedback never shows a pair that contradicts the decision.\n */\nfunction findDistinguishingPrecision(score: number, threshold: number | undefined): number {\n  const minimumPrecision = 6;\n  if (threshold === undefined || score === threshold) return minimumPrecision;\n  for (let precision = minimumPrecision; precision < 17; precision++) {\n    if (formatScore(score, precision) !== formatScore(threshold, precision)) return precision;\n  }\n  return 17;\n}\n\nfunction formatScore(score: number, precision: number): string {\n  return Number(score.toPrecision(precision)).toString();\n}\n\nfunction readPredictions(\n  submissionText: string,\n  submissionFilePath: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Map<string, string> | string {\n  let records: ReturnType<typeof parseCsvRecords>;\n  try {\n    records = parseCsvRecords(submissionText);\n  } catch (error) {\n    const detail =\n      error instanceof CsvParseError ? `${error.lineNumber}行目付近: ${CSV_PARSE_ERROR_MESSAGES[error.reason]}` : '';\n    return `\\`${submissionFilePath}\\`をCSVとして読み込めませんでした。${detail}`;\n  }\n  const columns = findColumns(records[0]?.cells ?? [], options);\n  if (!columns) {\n    return `\\`${submissionFilePath}\\`の1行目には\\`${options.idColumn}\\`列と\\`${options.targetColumn}\\`列のヘッダーが必要です。`;\n  }\n\n  const predictions = new Map<string, string>();\n  const problems: string[] = [];\n  let problemCount = 0;\n  const reportProblem = (problem: string): void => {\n    problemCount++;\n    if (problems.length < MAX_LISTED_IDS) problems.push(problem);\n  };\n  for (const { cells: row, lineNumber } of records.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const prediction = row[columns.targetIndex] ?? '';\n    if (!answer.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行は評価対象ではありません`);\n    } else if (predictions.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行が重複しています`);\n    } else {\n      const validationError = options.metric.validatePrediction?.(prediction);\n      if (validationError) {\n        reportProblem(`${lineNumber}行目: \\`${options.targetColumn}\\`の値\\`${clip(prediction)}\\`は${validationError}`);\n      }\n      predictions.set(id, prediction);\n    }\n  }\n  const missingIds = [...answer.keys()].filter((id) => !predictions.has(id));\n  if (missingIds.length > 0) {\n    // Missing rows are the most actionable problem, so they must survive the listing cap.\n    problemCount++;\n    problems.unshift(\n      `\\`${options.idColumn}\\`が${listIds(missingIds)}の行がありません（不足 ${missingIds.length.toLocaleString('en-US')}件）`\n    );\n  }\n  if (problemCount > 0) {\n    return `\\`${submissionFilePath}\\`の内容に問題があります。\n\n${problems.map((problem) => `- ${problem}`).join('\\n')}${problemCount > problems.length ? `\\n- ほか${(problemCount - problems.length).toLocaleString('en-US')}件` : ''}\n`;\n  }\n  return predictions;\n}\n\nasync function readAnswer(answerFilePath: string, options: EvaluationJudgePresetOptions): Promise<Map<string, string>> {\n  const rows = parseCsv(await fs.readFile(answerFilePath, 'utf8'));\n  const columns = findColumns(rows[0] ?? [], options);\n  if (!columns) {\n    throw new Error(`answer file must have ${options.idColumn} and ${options.targetColumn} columns: ${answerFilePath}`);\n  }\n  const answer = new Map<string, string>();\n  for (const row of rows.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const value = row[columns.targetIndex] ?? '';\n    if (!id) throw new Error(`empty ${options.idColumn} in answer file: ${answerFilePath}`);\n    if (answer.has(id)) throw new Error(`duplicate ${options.idColumn} in answer file: ${id}`);\n    const validationError = options.metric.validatePrediction?.(value);\n    if (validationError)\n      throw new Error(`invalid ${options.targetColumn} in answer file for ${id}: ${validationError}`);\n    answer.set(id, value);\n  }\n  if (answer.size === 0) throw new Error(`answer file has no rows: ${answerFilePath}`);\n  return answer;\n}\n\nfunction findColumns(\n  header: readonly string[],\n  options: EvaluationJudgePresetOptions\n): { idIndex: number; targetIndex: number } | undefined {\n  const names = header.map((name) => name.trim());\n  const idIndex = names.indexOf(options.idColumn);\n  const targetIndex = names.indexOf(options.targetColumn);\n  return idIndex === -1 || targetIndex === -1 ? undefined : { idIndex, targetIndex };\n}\n\nfunction isBlankRow(row: readonly string[]): boolean {\n  return row.every((cell) => cell.trim() === '');\n}\n\nfunction listIds(ids: readonly string[]): string {\n  const listed = ids.slice(0, MAX_LISTED_IDS).map((id) => `\\`${clip(id)}\\``);\n  return ids.length > MAX_LISTED_IDS ? `${listed.join(', ')}, ...` : listed.join(', ');\n}\n\nfunction clip(value: string): string {\n  // Control characters and backticks would break the markdown list item and inline code the value is echoed in.\n  const printable = value.replaceAll(/[\\p{Cc}]/gu, ' ').replaceAll('`', \"'\");\n  if (printable.trim() === '') return '（空）';\n  return printable.length > MAX_ECHOED_CELL_LENGTH ? `${printable.slice(0, MAX_ECHOED_CELL_LENGTH)}…` : printable;\n}\n\nfunction matchesExpectedResult(resolvedCwd: ResolvedCwd, result: Pick<TestCaseResult, 'decisionCode'>): boolean {\n  return (result.decisionCode === DecisionCode.ACCEPTED) === (resolvedCwd.expectedResult === 'accepted');\n}\n\nfunction validateFiniteNumber(prediction: string): string | undefined {\n  return Number.isFinite(Number(prediction)) && prediction.trim() !== '' ? undefined : '数値ではありません';\n}\n\nfunction mean(values: readonly number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length;\n}\n"],"mappings":"yqBAoBA,MAKM,EAAgE,CACpE,mBAAoB,qBACpB,iBAAkB,4BAClB,wBAAyB,yBAC3B,EAkCa,EAAoB,CAC/B,MAAO,CACL,MAAO,QACP,eAAgB,GAChB,mBAAqB,GACnB,EAAqB,CAAU,IAAM,OAAO,CAAU,EAAI,EAAI,QAAU,IAAA,IAC1E,SAAU,EAAgC,IACxC,KAAK,KACH,EACE,EAAY,KACT,EAAY,KAAW,KAAK,MAAM,OAAO,CAAU,CAAC,EAAI,KAAK,MAAM,OAAO,EAAQ,EAAM,CAAC,IAAM,CAClG,CACF,CACF,CACJ,EACA,KAAM,CACJ,MAAO,OACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,KAAK,KAAK,EAAK,EAAY,KAAK,EAAY,KAAW,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,IAAM,CAAC,CAAC,CAAC,CAC9G,EACA,IAAK,CACH,MAAO,MACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAU,KAAK,IAAI,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,CAAC,CAAC,CAAC,CACtG,EACA,SAAU,CACR,MAAO,WACP,eAAgB,GAChB,mBAAqB,GAAwB,EAAW,KAAK,IAAM,GAAK,MAAQ,IAAA,GAChF,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAW,IAAW,KAAK,IAAM,EAAQ,EAAM,EAAE,KAAK,EAAU,CAAC,CACvG,CACF,EAwBA,eAAsB,EAAsB,EAAoB,EAAsD,CACpH,IAAM,EAAOA,EAAAA,UAAU,QAAQ,IAAI,EAC7B,CAAE,OAAM,eAAgB,MAAMC,EAAAA,YAAY,EAAY,EAAK,GAAG,EAEpE,GAAI,EAAa,CACf,IAAM,EAAc,EAAK,KAAM,GAAQ,EAAI,iBAAmB,UAAU,EACxE,GAAI,EAEE,IAAA,EAAC,MAD8BC,EAAAA,yBAAyB,EAAY,EAAa,EAAK,MAAM,EAAA,CACtE,OAAQ,CAChC,QAAQ,SAAW,EACnB,MACF,OAEA,EAAA,iBAAiB,CACf,wDACA,GACA,4GACF,CAAC,CAEL,CAEA,IAAM,EAAS,MAAM,EAAWC,EAAAA,QAAK,KAAK,EAAY,EAAQ,cAAc,EAAG,CAAO,EACtF,IAAK,IAAM,KAAe,EAAM,CAC1B,GAAa,EAAA,oBAAoB,EAAY,CAAW,EAC5D,IAAM,EAAS,MAAM,EAAgB,EAAY,EAAY,IAAK,EAAQ,CAAO,EACjF,EAAA,oBAAoB,CAAE,WAAY,aAAc,GAAG,CAAO,CAAC,EACvD,GAAe,CAAC,EAAsB,EAAa,CAAM,IAC3D,QAAQ,SAAW,EACnB,EAAA,mCAAmC,EAAY,CAAW,EAE9D,CACF,CAEA,eAAe,EACb,EACA,EACA,EACA,EAC6C,CAC7C,IAAM,EAAuB,MAAMC,EAAAA,sBAAsB,EAAK,MAAMC,EAAAA,+BAA+B,CAAU,CAAC,EAC9G,GAAI,EAAsB,OAAO,EAEjC,IAAM,EAAqB,EAAQ,oBAAsB,iBACrD,EACJ,GAAI,CACF,IAAM,EAAeF,EAAAA,QAAK,KAAK,EAAK,CAAkB,EAEtD,GAAI,CAAE,MAAMG,EAAAA,2BAA2B,EAAK,CAAY,EAAI,MAAU,MAAM,aAAa,EACzF,EAAiB,MAAMC,EAAAA,QAAG,SAAS,EAAc,MAAM,CACzD,MAAQ,CACN,MAAO,CACL,aAAcC,EAAAA,aAAa,uCAC3B,iBAAkB,KAAK,EAAmB,4BAA4B,EAAmB,kBAC3F,CACF,CAEA,IAAM,EAAc,EAAgB,EAAgB,EAAoB,EAAQ,CAAO,EACvF,GAAI,OAAO,GAAgB,SACzB,MAAO,CAAE,aAAcA,EAAAA,aAAa,aAAc,iBAAkB,CAAY,EAGlF,IAAM,EAAe,CAAC,GAAG,EAAO,OAAO,CAAC,EAClC,EAAQ,EAAQ,OAAO,QAC3B,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,IAAK,GAAO,EAAY,IAAI,CAAE,CAAW,EAC5D,CACF,EACA,GAAI,CAAC,OAAO,SAAS,CAAK,EACxB,MAAO,CACL,aAAcA,EAAAA,aAAa,aAC3B,iBAAkB,aAAa,EAAQ,OAAO,MAAM,0BACtD,EAEF,IAAM,EACJ,EAAQ,kBAAoB,IAAA,KAC3B,EAAQ,OAAO,eAAiB,GAAS,EAAQ,gBAAkB,GAAS,EAAQ,iBACvF,MAAO,CACL,aAAc,EAAaA,EAAAA,aAAa,SAAWA,EAAAA,aAAa,aAChE,QACA,WAAY,EAAQ,OAAO,MAC3B,iBAAkB,EAAmB,EAAO,EAAa,OAAQ,EAAY,CAAO,CACtF,CACF,CAEA,SAAS,EACP,EACA,EACA,EACA,EACQ,CACR,GAAM,CAAE,QAAO,kBAAmB,EAAQ,OACpC,EAAY,EAA4B,EAAO,EAAQ,eAAe,EACtE,EACJ,EAAQ,kBAAoB,IAAA,GACxB,GACA,WAAW,EAAM,GAAG,EAAiB,IAAM,IAAI,GAAG,EAAY,EAAQ,gBAAiB,CAAS,EAAE,GAAG,EAAa,KAAO,MAAM,GACrI,MAAO;;IAEL,EAAM,KAAK,EAAY,EAAO,CAAS,EAAE;;QAErC,EAAS,eAAe,OAAO,EAAE,GAAG,EAAe;CAE3D,CAMA,SAAS,EAA4B,EAAe,EAAuC,CAEzF,GAAI,IAAc,IAAA,IAAa,IAAU,EAAW,MAAO,GAC3D,IAAK,IAAI,EAAY,EAAkB,EAAY,GAAI,IACrD,GAAI,EAAY,EAAO,CAAS,IAAM,EAAY,EAAW,CAAS,EAAG,OAAO,EAElF,MAAO,GACT,CAEA,SAAS,EAAY,EAAe,EAA2B,CAC7D,OAAO,OAAO,EAAM,YAAY,CAAS,CAAC,CAAC,CAAC,SAAS,CACvD,CAEA,SAAS,EACP,EACA,EACA,EACA,EAC8B,CAC9B,IAAI,EACJ,GAAI,CACF,EAAUC,EAAAA,gBAAgB,CAAc,CAC1C,OAAS,EAAO,CAGd,MAAO,KAAK,EAAmB,sBAD7B,aAAiBC,EAAAA,cAAgB,GAAG,EAAM,WAAW,QAAQ,EAAyB,EAAM,UAAY,IAE5G,CACA,IAAM,EAAU,EAAY,EAAQ,EAAE,EAAE,OAAS,CAAC,EAAG,CAAO,EAC5D,GAAI,CAAC,EACH,MAAO,KAAK,EAAmB,YAAY,EAAQ,SAAS,QAAQ,EAAQ,aAAa,gBAG3F,IAAM,EAAc,IAAI,IAClB,EAAqB,CAAC,EACxB,EAAe,EACb,EAAiB,GAA0B,CAC/C,IACI,EAAS,OAAS,IAAgB,EAAS,KAAK,CAAO,CAC7D,EACA,IAAK,GAAM,CAAE,MAAO,EAAK,gBAAgB,EAAQ,MAAM,CAAC,EAAG,CACzD,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAa,EAAI,EAAQ,cAAgB,GAC/C,GAAI,CAAC,EAAO,IAAI,CAAE,EAChB,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,iBAAiB,OACjF,GAAI,EAAY,IAAI,CAAE,EAC3B,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,aAAa,MAC7E,CACL,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAU,EAClE,GACF,EAAc,GAAG,EAAW,QAAQ,EAAQ,aAAa,QAAQ,EAAK,CAAU,EAAE,KAAK,GAAiB,EAE1G,EAAY,IAAI,EAAI,CAAU,CAChC,CACF,CACA,IAAM,EAAa,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,OAAQ,GAAO,CAAC,EAAY,IAAI,CAAE,CAAC,EAczE,OAbI,EAAW,OAAS,IAEtB,IACA,EAAS,QACP,KAAK,EAAQ,SAAS,KAAK,EAAQ,CAAU,EAAE,cAAc,EAAW,OAAO,eAAe,OAAO,EAAE,GACzG,GAEE,EAAe,EACV,KAAK,EAAmB;;EAEjC,EAAS,IAAK,GAAY,KAAK,GAAS,CAAC,CAAC,KAAK;CAAI,IAAI,EAAe,EAAS,OAAS,UAAU,EAAe,EAAS,OAAA,CAAQ,eAAe,OAAO,EAAE,GAAK,GAAG;EAG3J,CACT,CAEA,eAAe,EAAW,EAAwB,EAAqE,CACrH,IAAM,EAAOC,EAAAA,SAAS,MAAMJ,EAAAA,QAAG,SAAS,EAAgB,MAAM,CAAC,EACzD,EAAU,EAAY,EAAK,IAAM,CAAC,EAAG,CAAO,EAClD,GAAI,CAAC,EACH,MAAU,MAAM,yBAAyB,EAAQ,SAAS,OAAO,EAAQ,aAAa,YAAY,GAAgB,EAEpH,IAAM,EAAS,IAAI,IACnB,IAAK,IAAM,KAAO,EAAK,MAAM,CAAC,EAAG,CAC/B,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAQ,EAAI,EAAQ,cAAgB,GAC1C,GAAI,CAAC,EAAI,MAAU,MAAM,SAAS,EAAQ,SAAS,mBAAmB,GAAgB,EACtF,GAAI,EAAO,IAAI,CAAE,EAAG,MAAU,MAAM,aAAa,EAAQ,SAAS,mBAAmB,GAAI,EACzF,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAK,EACjE,GAAI,EACF,MAAU,MAAM,WAAW,EAAQ,aAAa,sBAAsB,EAAG,IAAI,GAAiB,EAChG,EAAO,IAAI,EAAI,CAAK,CACtB,CACA,GAAI,EAAO,OAAS,EAAG,MAAU,MAAM,4BAA4B,GAAgB,EACnF,OAAO,CACT,CAEA,SAAS,EACP,EACA,EACsD,CACtD,IAAM,EAAQ,EAAO,IAAK,GAAS,EAAK,KAAK,CAAC,EACxC,EAAU,EAAM,QAAQ,EAAQ,QAAQ,EACxC,EAAc,EAAM,QAAQ,EAAQ,YAAY,EACtD,OAAO,IAAY,IAAM,IAAgB,GAAK,IAAA,GAAY,CAAE,UAAS,aAAY,CACnF,CAEA,SAAS,EAAW,EAAiC,CACnD,OAAO,EAAI,MAAO,GAAS,EAAK,KAAK,IAAM,EAAE,CAC/C,CAEA,SAAS,EAAQ,EAAgC,CAC/C,IAAM,EAAS,EAAI,MAAM,EAAG,EAAc,CAAC,CAAC,IAAK,GAAO,KAAK,EAAK,CAAE,EAAE,GAAG,EACzE,OAAO,EAAI,OAAS,GAAiB,GAAG,EAAO,KAAK,IAAI,EAAE,OAAS,EAAO,KAAK,IAAI,CACrF,CAEA,SAAS,EAAK,EAAuB,CAEnC,IAAM,EAAY,EAAM,WAAW,aAAc,GAAG,CAAC,CAAC,WAAW,IAAK,GAAG,EAEzE,OADI,EAAU,KAAK,IAAM,GAAW,MAC7B,EAAU,OAAS,GAAyB,GAAG,EAAU,MAAM,EAAG,EAAsB,EAAE,GAAK,CACxG,CAEA,SAAS,EAAsB,EAA0B,EAAuD,CAC9G,OAAQ,EAAO,eAAiBC,EAAAA,aAAa,WAAe,EAAY,iBAAmB,WAC7F,CAEA,SAAS,EAAqB,EAAwC,CACpE,OAAO,OAAO,SAAS,OAAO,CAAU,CAAC,GAAK,EAAW,KAAK,IAAM,GAAK,IAAA,GAAY,WACvF,CAEA,SAAS,EAAK,EAAmC,CAC/C,OAAO,EAAO,QAAQ,EAAK,IAAU,EAAM,EAAO,CAAC,EAAI,EAAO,MAChE"}
+{"version":3,"file":"evaluation.cjs","names":["parseArgs","resolveCwds","passesIsolationCheckInDebugMode","path","judgeByStaticAnalysis","readProblemMarkdownFrontMatter","isSafeSubmissionOutputPath","fs","DecisionCode","parseCsvRecords","CsvParseError","parseCsv"],"sources":["../../src/presets/evaluation.ts"],"sourcesContent":["import fs from 'node:fs/promises';\nimport path from 'node:path';\n\nimport { judgeByStaticAnalysis } from '../helpers/judgeByStaticAnalysis.js';\nimport { parseArgs } from '../helpers/parseArgs.js';\nimport { CsvParseError, type CsvParseErrorReason, parseCsv, parseCsvRecords } from '../helpers/parseCsv.js';\nimport { printTestCaseResult } from '../helpers/printTestCaseResult.js';\nimport { isSafeSubmissionOutputPath } from '../helpers/readOutputFiles.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport { passesIsolationCheckInDebugMode } from '../helpers/runIsolationCheckInDebugMode.js';\nimport {\n  printDebugCwdBanner,\n  printDebugExpectationFailureBanner,\n  resolveCwds,\n  type ResolvedCwd,\n} from '../helpers/resolveCwds.js';\nimport { DecisionCode } from '../types/decisionCode.js';\nimport type { TestCaseResult } from '../types/testCaseResult.js';\n\nconst DEFAULT_SUBMISSION_FILE_PATH = 'submission.csv';\nconst TEST_CASE_ID = 'evaluation';\nconst MAX_LISTED_IDS = 10;\n// Learner-controlled cells are echoed into the feedback, so clip them to keep the result line small.\nconst MAX_ECHOED_CELL_LENGTH = 50;\nconst CSV_PARSE_ERROR_MESSAGES: Record<CsvParseErrorReason, string> = {\n  unterminated_quote: '引用符（`\"`）が閉じられていません',\n  unexpected_quote: '引用符（`\"`）は値全体を囲む位置にしか書けません',\n  text_after_quoted_field: '閉じ引用符の後に区切り文字以外の文字があります',\n};\n\nexport interface EvaluationMetric {\n  /** Label shown with the score, e.g. `RMSLE`. */\n  label: string;\n  isHigherBetter: boolean;\n  /**\n   * Returns an error message when a predicted value cannot be scored (e.g. it is not a number).\n   * The message is appended to the row's line number and the offending value, so phrase it as a\n   * predicate such as `数値ではありません`. It also validates the ground truth when it is loaded.\n   */\n  validatePrediction?: (prediction: string) => string | undefined;\n  /** Computes the score from predictions and answers aligned by id. Both are raw CSV cell strings. */\n  compute: (predictions: readonly string[], answers: readonly string[]) => number;\n}\n\nexport interface EvaluationJudgePresetOptions {\n  /** CSV file holding the hidden ground truth, relative to the problem directory. */\n  answerFilePath: string;\n  /** CSV file the submission must contain. Defaults to `submission.csv`. */\n  submissionFilePath?: string;\n  /** Column that identifies each row in both CSV files. */\n  idColumn: string;\n  /** Column holding the answer in the ground truth and the prediction in the submission. */\n  targetColumn: string;\n  metric: EvaluationMetric;\n  /**\n   * Score required for acceptance: the submission is accepted when its score is at least this value\n   * for a higher-is-better metric and at most this value otherwise. Every valid submission is\n   * accepted when omitted.\n   */\n  acceptableScore?: number;\n}\n\nexport const evaluationMetrics = {\n  rmsle: {\n    label: 'RMSLE',\n    isHigherBetter: false,\n    validatePrediction: (prediction: string) =>\n      validateFiniteNumber(prediction) ?? (Number(prediction) < 0 ? '負の値です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(\n        mean(\n          predictions.map(\n            (prediction, index) => (Math.log1p(Number(prediction)) - Math.log1p(Number(answers[index]))) ** 2\n          )\n        )\n      ),\n  },\n  rmse: {\n    label: 'RMSE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(mean(predictions.map((prediction, index) => (Number(prediction) - Number(answers[index])) ** 2))),\n  },\n  mae: {\n    label: 'MAE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => Math.abs(Number(prediction) - Number(answers[index])))),\n  },\n  accuracy: {\n    label: 'Accuracy',\n    isHigherBetter: true,\n    validatePrediction: (prediction: string) => (prediction.trim() === '' ? '空です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => (prediction.trim() === answers[index]?.trim() ? 1 : 0))),\n  },\n} as const satisfies Record<string, EvaluationMetric>;\n\n/**\n * A preset for Kaggle-style model evaluation problems: the submission is a CSV of predictions for a\n * hidden test set, and the judge reports the metric value as `score` instead of comparing outputs.\n * The problem directory keeps the ground truth CSV, so it is never shipped to learners.\n *\n * @example\n * Create `judge.ts`:\n * ```ts\n * import { evaluationJudgePreset, evaluationMetrics } from '@exercode/problem-utils/presets/evaluation';\n *\n * await evaluationJudgePreset(import.meta.dirname, {\n *   answerFilePath: 'evaluation/answer.csv',\n *   idColumn: 'Id',\n *   targetColumn: 'TradePrice',\n *   metric: evaluationMetrics.rmsle,\n *   acceptableScore: 0.5,\n * });\n * ```\n *\n * Run `bun judge.ts model_answers/default` to judge one submission, or `bun judge.ts` to judge every\n * `<problemDir>/model_answers/*` and `<problemDir>/model_answers.fails/*` directory for debugging.\n */\nexport async function evaluationJudgePreset(problemDir: string, options: EvaluationJudgePresetOptions): Promise<void> {\n  const args = parseArgs(process.argv);\n  const { cwds, isDebugMode } = await resolveCwds(problemDir, args.cwd);\n\n  if (isDebugMode && !(await passesIsolationCheckInDebugMode(problemDir, cwds, args.params))) {\n    process.exitCode = 1;\n    return;\n  }\n\n  const answer = await readAnswer(path.join(problemDir, options.answerFilePath), options);\n  for (const resolvedCwd of cwds) {\n    if (isDebugMode) printDebugCwdBanner(problemDir, resolvedCwd);\n    const result = await judgeSubmission(problemDir, resolvedCwd.cwd, answer, options);\n    printTestCaseResult({ testCaseId: TEST_CASE_ID, ...result });\n    if (isDebugMode && !matchesExpectedResult(resolvedCwd, result)) {\n      process.exitCode = 1;\n      printDebugExpectationFailureBanner(problemDir, resolvedCwd);\n    }\n  }\n}\n\nasync function judgeSubmission(\n  problemDir: string,\n  cwd: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Promise<Omit<TestCaseResult, 'testCaseId'>> {\n  const staticAnalysisResult = await judgeByStaticAnalysis(cwd, await readProblemMarkdownFrontMatter(problemDir));\n  if (staticAnalysisResult) return staticAnalysisResult;\n\n  const submissionFilePath = options.submissionFilePath ?? DEFAULT_SUBMISSION_FILE_PATH;\n  let submissionText: string;\n  try {\n    const resolvedPath = path.join(cwd, submissionFilePath);\n    // The harness reads the file as the trusted user, so refuse symlinks escaping the submission.\n    if (!(await isSafeSubmissionOutputPath(cwd, resolvedPath))) throw new Error('unsafe path');\n    submissionText = await fs.readFile(resolvedPath, 'utf8');\n  } catch {\n    return {\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      feedbackMarkdown: `\\`${submissionFilePath}\\`が見つかりません。予測結果のCSVファイルを\\`${submissionFilePath}\\`という名前で提出してください。`,\n    };\n  }\n\n  const predictions = readPredictions(submissionText, submissionFilePath, answer, options);\n  if (typeof predictions === 'string') {\n    return { decisionCode: DecisionCode.WRONG_ANSWER, feedbackMarkdown: predictions };\n  }\n\n  const answerValues = [...answer.values()];\n  const score = options.metric.compute(\n    [...answer.keys()].map((id) => predictions.get(id) as string),\n    answerValues\n  );\n  if (!Number.isFinite(score)) {\n    return {\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      feedbackMarkdown: `予測値が極端なため、${options.metric.label}を計算できませんでした。予測値を見直してください。`,\n    };\n  }\n  const isAccepted =\n    options.acceptableScore === undefined ||\n    (options.metric.isHigherBetter ? score >= options.acceptableScore : score <= options.acceptableScore);\n  return {\n    decisionCode: isAccepted ? DecisionCode.ACCEPTED : DecisionCode.WRONG_ANSWER,\n    score,\n    scoreLabel: options.metric.label,\n    feedbackMarkdown: buildScoreMarkdown(score, answerValues.length, isAccepted, options),\n  };\n}\n\nfunction buildScoreMarkdown(\n  score: number,\n  rowCount: number,\n  isAccepted: boolean,\n  options: EvaluationJudgePresetOptions\n): string {\n  const { label, isHigherBetter } = options.metric;\n  const precision = findDistinguishingPrecision(score, options.acceptableScore);\n  const acceptanceLine =\n    options.acceptableScore === undefined\n      ? ''\n      : `\\n合格基準: ${label} ${isHigherBetter ? '≧' : '≦'} ${formatScore(options.acceptableScore, precision)}（${isAccepted ? '達成' : '未達成'}）`;\n  return `| 指標 | スコア |\n| ---- | ------ |\n| ${label} | ${formatScore(score, precision)} |\n\n評価件数: ${rowCount.toLocaleString('en-US')}件${acceptanceLine}\n`;\n}\n\n/**\n * Finds the number of significant digits at which the score and the threshold render differently\n * (or stop changing), so the feedback never shows a pair that contradicts the decision.\n */\nfunction findDistinguishingPrecision(score: number, threshold: number | undefined): number {\n  const minimumPrecision = 6;\n  if (threshold === undefined || score === threshold) return minimumPrecision;\n  for (let precision = minimumPrecision; precision < 17; precision++) {\n    if (formatScore(score, precision) !== formatScore(threshold, precision)) return precision;\n  }\n  return 17;\n}\n\nfunction formatScore(score: number, precision: number): string {\n  return Number(score.toPrecision(precision)).toString();\n}\n\nfunction readPredictions(\n  submissionText: string,\n  submissionFilePath: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Map<string, string> | string {\n  let records: ReturnType<typeof parseCsvRecords>;\n  try {\n    records = parseCsvRecords(submissionText);\n  } catch (error) {\n    const detail =\n      error instanceof CsvParseError ? `${error.lineNumber}行目付近: ${CSV_PARSE_ERROR_MESSAGES[error.reason]}` : '';\n    return `\\`${submissionFilePath}\\`をCSVとして読み込めませんでした。${detail}`;\n  }\n  const columns = findColumns(records[0]?.cells ?? [], options);\n  if (!columns) {\n    return `\\`${submissionFilePath}\\`の1行目には\\`${options.idColumn}\\`列と\\`${options.targetColumn}\\`列のヘッダーが必要です。`;\n  }\n\n  const predictions = new Map<string, string>();\n  const problems: string[] = [];\n  let problemCount = 0;\n  const reportProblem = (problem: string): void => {\n    problemCount++;\n    if (problems.length < MAX_LISTED_IDS) problems.push(problem);\n  };\n  for (const { cells: row, lineNumber } of records.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const prediction = row[columns.targetIndex] ?? '';\n    if (!answer.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行は評価対象ではありません`);\n    } else if (predictions.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行が重複しています`);\n    } else {\n      const validationError = options.metric.validatePrediction?.(prediction);\n      if (validationError) {\n        reportProblem(`${lineNumber}行目: \\`${options.targetColumn}\\`の値\\`${clip(prediction)}\\`は${validationError}`);\n      }\n      predictions.set(id, prediction);\n    }\n  }\n  const missingIds = [...answer.keys()].filter((id) => !predictions.has(id));\n  if (missingIds.length > 0) {\n    // Missing rows are the most actionable problem, so they must survive the listing cap.\n    problemCount++;\n    problems.unshift(\n      `\\`${options.idColumn}\\`が${listIds(missingIds)}の行がありません（不足 ${missingIds.length.toLocaleString('en-US')}件）`\n    );\n  }\n  if (problemCount > 0) {\n    return `\\`${submissionFilePath}\\`の内容に問題があります。\n\n${problems.map((problem) => `- ${problem}`).join('\\n')}${problemCount > problems.length ? `\\n- ほか${(problemCount - problems.length).toLocaleString('en-US')}件` : ''}\n`;\n  }\n  return predictions;\n}\n\nasync function readAnswer(answerFilePath: string, options: EvaluationJudgePresetOptions): Promise<Map<string, string>> {\n  const rows = parseCsv(await fs.readFile(answerFilePath, 'utf8'));\n  const columns = findColumns(rows[0] ?? [], options);\n  if (!columns) {\n    throw new Error(`answer file must have ${options.idColumn} and ${options.targetColumn} columns: ${answerFilePath}`);\n  }\n  const answer = new Map<string, string>();\n  for (const row of rows.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const value = row[columns.targetIndex] ?? '';\n    if (!id) throw new Error(`empty ${options.idColumn} in answer file: ${answerFilePath}`);\n    if (answer.has(id)) throw new Error(`duplicate ${options.idColumn} in answer file: ${id}`);\n    const validationError = options.metric.validatePrediction?.(value);\n    if (validationError)\n      throw new Error(`invalid ${options.targetColumn} in answer file for ${id}: ${validationError}`);\n    answer.set(id, value);\n  }\n  if (answer.size === 0) throw new Error(`answer file has no rows: ${answerFilePath}`);\n  return answer;\n}\n\nfunction findColumns(\n  header: readonly string[],\n  options: EvaluationJudgePresetOptions\n): { idIndex: number; targetIndex: number } | undefined {\n  const names = header.map((name) => name.trim());\n  const idIndex = names.indexOf(options.idColumn);\n  const targetIndex = names.indexOf(options.targetColumn);\n  return idIndex === -1 || targetIndex === -1 ? undefined : { idIndex, targetIndex };\n}\n\nfunction isBlankRow(row: readonly string[]): boolean {\n  return row.every((cell) => cell.trim() === '');\n}\n\nfunction listIds(ids: readonly string[]): string {\n  const listed = ids.slice(0, MAX_LISTED_IDS).map((id) => `\\`${clip(id)}\\``);\n  return ids.length > MAX_LISTED_IDS ? `${listed.join(', ')}, ...` : listed.join(', ');\n}\n\nfunction clip(value: string): string {\n  // Control characters and backticks would break the markdown list item and inline code the value is echoed in.\n  const printable = value.replaceAll(/[\\p{Cc}]/gu, ' ').replaceAll('`', \"'\");\n  if (printable.trim() === '') return '（空）';\n  return printable.length > MAX_ECHOED_CELL_LENGTH ? `${printable.slice(0, MAX_ECHOED_CELL_LENGTH)}…` : printable;\n}\n\nfunction matchesExpectedResult(resolvedCwd: ResolvedCwd, result: Pick<TestCaseResult, 'decisionCode'>): boolean {\n  return (result.decisionCode === DecisionCode.ACCEPTED) === (resolvedCwd.expectedResult === 'accepted');\n}\n\nfunction validateFiniteNumber(prediction: string): string | undefined {\n  return Number.isFinite(Number(prediction)) && prediction.trim() !== '' ? undefined : '数値ではありません';\n}\n\nfunction mean(values: readonly number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length;\n}\n"],"mappings":"goBAmBA,MAKM,EAAgE,CACpE,mBAAoB,qBACpB,iBAAkB,4BAClB,wBAAyB,yBAC3B,EAkCa,EAAoB,CAC/B,MAAO,CACL,MAAO,QACP,eAAgB,GAChB,mBAAqB,GACnB,EAAqB,CAAU,IAAM,OAAO,CAAU,EAAI,EAAI,QAAU,IAAA,IAC1E,SAAU,EAAgC,IACxC,KAAK,KACH,EACE,EAAY,KACT,EAAY,KAAW,KAAK,MAAM,OAAO,CAAU,CAAC,EAAI,KAAK,MAAM,OAAO,EAAQ,EAAM,CAAC,IAAM,CAClG,CACF,CACF,CACJ,EACA,KAAM,CACJ,MAAO,OACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,KAAK,KAAK,EAAK,EAAY,KAAK,EAAY,KAAW,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,IAAM,CAAC,CAAC,CAAC,CAC9G,EACA,IAAK,CACH,MAAO,MACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAU,KAAK,IAAI,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,CAAC,CAAC,CAAC,CACtG,EACA,SAAU,CACR,MAAO,WACP,eAAgB,GAChB,mBAAqB,GAAwB,EAAW,KAAK,IAAM,GAAK,MAAQ,IAAA,GAChF,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAW,IAAW,KAAK,IAAM,EAAQ,EAAM,EAAE,KAAK,EAAU,CAAC,CACvG,CACF,EAwBA,eAAsB,EAAsB,EAAoB,EAAsD,CACpH,IAAM,EAAOA,EAAAA,UAAU,QAAQ,IAAI,EAC7B,CAAE,OAAM,eAAgB,MAAMC,EAAAA,YAAY,EAAY,EAAK,GAAG,EAEpE,GAAI,GAAe,CAAE,MAAMC,EAAAA,gCAAgC,EAAY,EAAM,EAAK,MAAM,EAAI,CAC1F,QAAQ,SAAW,EACnB,MACF,CAEA,IAAM,EAAS,MAAM,EAAWC,EAAAA,QAAK,KAAK,EAAY,EAAQ,cAAc,EAAG,CAAO,EACtF,IAAK,IAAM,KAAe,EAAM,CAC1B,GAAa,EAAA,oBAAoB,EAAY,CAAW,EAC5D,IAAM,EAAS,MAAM,EAAgB,EAAY,EAAY,IAAK,EAAQ,CAAO,EACjF,EAAA,oBAAoB,CAAE,WAAY,aAAc,GAAG,CAAO,CAAC,EACvD,GAAe,CAAC,EAAsB,EAAa,CAAM,IAC3D,QAAQ,SAAW,EACnB,EAAA,mCAAmC,EAAY,CAAW,EAE9D,CACF,CAEA,eAAe,EACb,EACA,EACA,EACA,EAC6C,CAC7C,IAAM,EAAuB,MAAMC,EAAAA,sBAAsB,EAAK,MAAMC,EAAAA,+BAA+B,CAAU,CAAC,EAC9G,GAAI,EAAsB,OAAO,EAEjC,IAAM,EAAqB,EAAQ,oBAAsB,iBACrD,EACJ,GAAI,CACF,IAAM,EAAeF,EAAAA,QAAK,KAAK,EAAK,CAAkB,EAEtD,GAAI,CAAE,MAAMG,EAAAA,2BAA2B,EAAK,CAAY,EAAI,MAAU,MAAM,aAAa,EACzF,EAAiB,MAAMC,EAAAA,QAAG,SAAS,EAAc,MAAM,CACzD,MAAQ,CACN,MAAO,CACL,aAAcC,EAAAA,aAAa,uCAC3B,iBAAkB,KAAK,EAAmB,4BAA4B,EAAmB,kBAC3F,CACF,CAEA,IAAM,EAAc,EAAgB,EAAgB,EAAoB,EAAQ,CAAO,EACvF,GAAI,OAAO,GAAgB,SACzB,MAAO,CAAE,aAAcA,EAAAA,aAAa,aAAc,iBAAkB,CAAY,EAGlF,IAAM,EAAe,CAAC,GAAG,EAAO,OAAO,CAAC,EAClC,EAAQ,EAAQ,OAAO,QAC3B,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,IAAK,GAAO,EAAY,IAAI,CAAE,CAAW,EAC5D,CACF,EACA,GAAI,CAAC,OAAO,SAAS,CAAK,EACxB,MAAO,CACL,aAAcA,EAAAA,aAAa,aAC3B,iBAAkB,aAAa,EAAQ,OAAO,MAAM,0BACtD,EAEF,IAAM,EACJ,EAAQ,kBAAoB,IAAA,KAC3B,EAAQ,OAAO,eAAiB,GAAS,EAAQ,gBAAkB,GAAS,EAAQ,iBACvF,MAAO,CACL,aAAc,EAAaA,EAAAA,aAAa,SAAWA,EAAAA,aAAa,aAChE,QACA,WAAY,EAAQ,OAAO,MAC3B,iBAAkB,EAAmB,EAAO,EAAa,OAAQ,EAAY,CAAO,CACtF,CACF,CAEA,SAAS,EACP,EACA,EACA,EACA,EACQ,CACR,GAAM,CAAE,QAAO,kBAAmB,EAAQ,OACpC,EAAY,EAA4B,EAAO,EAAQ,eAAe,EACtE,EACJ,EAAQ,kBAAoB,IAAA,GACxB,GACA,WAAW,EAAM,GAAG,EAAiB,IAAM,IAAI,GAAG,EAAY,EAAQ,gBAAiB,CAAS,EAAE,GAAG,EAAa,KAAO,MAAM,GACrI,MAAO;;IAEL,EAAM,KAAK,EAAY,EAAO,CAAS,EAAE;;QAErC,EAAS,eAAe,OAAO,EAAE,GAAG,EAAe;CAE3D,CAMA,SAAS,EAA4B,EAAe,EAAuC,CAEzF,GAAI,IAAc,IAAA,IAAa,IAAU,EAAW,MAAO,GAC3D,IAAK,IAAI,EAAY,EAAkB,EAAY,GAAI,IACrD,GAAI,EAAY,EAAO,CAAS,IAAM,EAAY,EAAW,CAAS,EAAG,OAAO,EAElF,MAAO,GACT,CAEA,SAAS,EAAY,EAAe,EAA2B,CAC7D,OAAO,OAAO,EAAM,YAAY,CAAS,CAAC,CAAC,CAAC,SAAS,CACvD,CAEA,SAAS,EACP,EACA,EACA,EACA,EAC8B,CAC9B,IAAI,EACJ,GAAI,CACF,EAAUC,EAAAA,gBAAgB,CAAc,CAC1C,OAAS,EAAO,CAGd,MAAO,KAAK,EAAmB,sBAD7B,aAAiBC,EAAAA,cAAgB,GAAG,EAAM,WAAW,QAAQ,EAAyB,EAAM,UAAY,IAE5G,CACA,IAAM,EAAU,EAAY,EAAQ,EAAE,EAAE,OAAS,CAAC,EAAG,CAAO,EAC5D,GAAI,CAAC,EACH,MAAO,KAAK,EAAmB,YAAY,EAAQ,SAAS,QAAQ,EAAQ,aAAa,gBAG3F,IAAM,EAAc,IAAI,IAClB,EAAqB,CAAC,EACxB,EAAe,EACb,EAAiB,GAA0B,CAC/C,IACI,EAAS,OAAS,IAAgB,EAAS,KAAK,CAAO,CAC7D,EACA,IAAK,GAAM,CAAE,MAAO,EAAK,gBAAgB,EAAQ,MAAM,CAAC,EAAG,CACzD,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAa,EAAI,EAAQ,cAAgB,GAC/C,GAAI,CAAC,EAAO,IAAI,CAAE,EAChB,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,iBAAiB,OACjF,GAAI,EAAY,IAAI,CAAE,EAC3B,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,aAAa,MAC7E,CACL,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAU,EAClE,GACF,EAAc,GAAG,EAAW,QAAQ,EAAQ,aAAa,QAAQ,EAAK,CAAU,EAAE,KAAK,GAAiB,EAE1G,EAAY,IAAI,EAAI,CAAU,CAChC,CACF,CACA,IAAM,EAAa,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,OAAQ,GAAO,CAAC,EAAY,IAAI,CAAE,CAAC,EAczE,OAbI,EAAW,OAAS,IAEtB,IACA,EAAS,QACP,KAAK,EAAQ,SAAS,KAAK,EAAQ,CAAU,EAAE,cAAc,EAAW,OAAO,eAAe,OAAO,EAAE,GACzG,GAEE,EAAe,EACV,KAAK,EAAmB;;EAEjC,EAAS,IAAK,GAAY,KAAK,GAAS,CAAC,CAAC,KAAK;CAAI,IAAI,EAAe,EAAS,OAAS,UAAU,EAAe,EAAS,OAAA,CAAQ,eAAe,OAAO,EAAE,GAAK,GAAG;EAG3J,CACT,CAEA,eAAe,EAAW,EAAwB,EAAqE,CACrH,IAAM,EAAOC,EAAAA,SAAS,MAAMJ,EAAAA,QAAG,SAAS,EAAgB,MAAM,CAAC,EACzD,EAAU,EAAY,EAAK,IAAM,CAAC,EAAG,CAAO,EAClD,GAAI,CAAC,EACH,MAAU,MAAM,yBAAyB,EAAQ,SAAS,OAAO,EAAQ,aAAa,YAAY,GAAgB,EAEpH,IAAM,EAAS,IAAI,IACnB,IAAK,IAAM,KAAO,EAAK,MAAM,CAAC,EAAG,CAC/B,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAQ,EAAI,EAAQ,cAAgB,GAC1C,GAAI,CAAC,EAAI,MAAU,MAAM,SAAS,EAAQ,SAAS,mBAAmB,GAAgB,EACtF,GAAI,EAAO,IAAI,CAAE,EAAG,MAAU,MAAM,aAAa,EAAQ,SAAS,mBAAmB,GAAI,EACzF,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAK,EACjE,GAAI,EACF,MAAU,MAAM,WAAW,EAAQ,aAAa,sBAAsB,EAAG,IAAI,GAAiB,EAChG,EAAO,IAAI,EAAI,CAAK,CACtB,CACA,GAAI,EAAO,OAAS,EAAG,MAAU,MAAM,4BAA4B,GAAgB,EACnF,OAAO,CACT,CAEA,SAAS,EACP,EACA,EACsD,CACtD,IAAM,EAAQ,EAAO,IAAK,GAAS,EAAK,KAAK,CAAC,EACxC,EAAU,EAAM,QAAQ,EAAQ,QAAQ,EACxC,EAAc,EAAM,QAAQ,EAAQ,YAAY,EACtD,OAAO,IAAY,IAAM,IAAgB,GAAK,IAAA,GAAY,CAAE,UAAS,aAAY,CACnF,CAEA,SAAS,EAAW,EAAiC,CACnD,OAAO,EAAI,MAAO,GAAS,EAAK,KAAK,IAAM,EAAE,CAC/C,CAEA,SAAS,EAAQ,EAAgC,CAC/C,IAAM,EAAS,EAAI,MAAM,EAAG,EAAc,CAAC,CAAC,IAAK,GAAO,KAAK,EAAK,CAAE,EAAE,GAAG,EACzE,OAAO,EAAI,OAAS,GAAiB,GAAG,EAAO,KAAK,IAAI,EAAE,OAAS,EAAO,KAAK,IAAI,CACrF,CAEA,SAAS,EAAK,EAAuB,CAEnC,IAAM,EAAY,EAAM,WAAW,aAAc,GAAG,CAAC,CAAC,WAAW,IAAK,GAAG,EAEzE,OADI,EAAU,KAAK,IAAM,GAAW,MAC7B,EAAU,OAAS,GAAyB,GAAG,EAAU,MAAM,EAAG,EAAsB,EAAE,GAAK,CACxG,CAEA,SAAS,EAAsB,EAA0B,EAAuD,CAC9G,OAAQ,EAAO,eAAiBC,EAAAA,aAAa,WAAe,EAAY,iBAAmB,WAC7F,CAEA,SAAS,EAAqB,EAAwC,CACpE,OAAO,OAAO,SAAS,OAAO,CAAU,CAAC,GAAK,EAAW,KAAK,IAAM,GAAK,IAAA,GAAY,WACvF,CAEA,SAAS,EAAK,EAAmC,CAC/C,OAAO,EAAO,QAAQ,EAAK,IAAU,EAAM,EAAO,CAAC,EAAI,EAAO,MAChE"}

package/dist/presets/evaluation.js

@@ -1,11 +1,11 @@
-import{parseArgs as e}from"../helpers/parseArgs.js";import{CsvParseError as t,parseCsv as n,parseCsvRecords as r}from"../helpers/parseCsv.js";import{printTestCaseResult as i}from"../helpers/printTestCaseResult.js";import{DecisionCode as a}from"../types/decisionCode.js";import{printDebugBanner as o}from"../helpers/printDebugBanner.js";import{checkProblemDirIsolation as s}from"../helpers/checkProblemDirIsolation.js";import{judgeByStaticAnalysis as c}from"../helpers/judgeByStaticAnalysis.js";import{isSafeSubmissionOutputPath as l}from"../helpers/readOutputFiles.js";import{readProblemMarkdownFrontMatter as u}from"../helpers/readProblemMarkdownFrontMatter.js";import{printDebugCwdBanner as d,printDebugExpectationFailureBanner as f,resolveCwds as p}from"../helpers/resolveCwds.js";import m from"node:fs/promises";import h from"node:path";const g={unterminated_quote:'引用符（`"`）が閉じられていません',unexpected_quote:'引用符（`"`）は値全体を囲む位置にしか書けません',text_after_quoted_field:`閉じ引用符の後に区切り文字以外の文字があります`},_={rmsle:{label:`RMSLE`,isHigherBetter:!1,validatePrediction:e=>A(e)??(Number(e)<0?`負の値です`:void 0),compute:(e,t)=>Math.sqrt(j(e.map((e,n)=>(Math.log1p(Number(e))-Math.log1p(Number(t[n])))**2)))},rmse:{label:`RMSE`,isHigherBetter:!1,validatePrediction:A,compute:(e,t)=>Math.sqrt(j(e.map((e,n)=>(Number(e)-Number(t[n]))**2)))},mae:{label:`MAE`,isHigherBetter:!1,validatePrediction:A,compute:(e,t)=>j(e.map((e,n)=>Math.abs(Number(e)-Number(t[n]))))},accuracy:{label:`Accuracy`,isHigherBetter:!0,validatePrediction:e=>e.trim()===``?`空です`:void 0,compute:(e,t)=>j(e.map((e,n)=>+(e.trim()===t[n]?.trim())))}};async function v(t,n){let r=e(process.argv),{cwds:a,isDebugMode:c}=await p(t,r.cwd);if(c){let e=a.find(e=>e.expectedResult===`accepted`);if(e){if(!(await s(t,e,r.params)).passed){process.exitCode=1;return}}else o([`[DEBUG MODE] isolated problem directory check skipped`,``,`No accepted model answer is available for checking that the copied judge still accepts a valid submission.`])}let l=await w(h.join(t,n.answerFilePath),n);for(let e of a){c&&d(t,e);let r=await y(t,e.cwd,l,n);i({testCaseId:`evaluation`,...r}),c&&!k(e,r)&&(process.exitCode=1,f(t,e))}}async function y(e,t,n,r){let i=await c(t,await u(e));if(i)return i;let o=r.submissionFilePath??`submission.csv`,s;try{let e=h.join(t,o);if(!await l(t,e))throw Error(`unsafe path`);s=await m.readFile(e,`utf8`)}catch{return{decisionCode:a.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,feedbackMarkdown:`\`${o}\`が見つかりません。予測結果のCSVファイルを\`${o}\`という名前で提出してください。`}}let d=C(s,o,n,r);if(typeof d==`string`)return{decisionCode:a.WRONG_ANSWER,feedbackMarkdown:d};let f=[...n.values()],p=r.metric.compute([...n.keys()].map(e=>d.get(e)),f);if(!Number.isFinite(p))return{decisionCode:a.WRONG_ANSWER,feedbackMarkdown:`予測値が極端なため、${r.metric.label}を計算できませんでした。予測値を見直してください。`};let g=r.acceptableScore===void 0||(r.metric.isHigherBetter?p>=r.acceptableScore:p<=r.acceptableScore);return{decisionCode:g?a.ACCEPTED:a.WRONG_ANSWER,score:p,scoreLabel:r.metric.label,feedbackMarkdown:b(p,f.length,g,r)}}function b(e,t,n,r){let{label:i,isHigherBetter:a}=r.metric,o=x(e,r.acceptableScore),s=r.acceptableScore===void 0?``:`\n合格基準: ${i} ${a?`≧`:`≦`} ${S(r.acceptableScore,o)}（${n?`達成`:`未達成`}）`;return`| 指標 | スコア |
+import{parseArgs as e}from"../helpers/parseArgs.js";import{CsvParseError as t,parseCsv as n,parseCsvRecords as r}from"../helpers/parseCsv.js";import{printTestCaseResult as i}from"../helpers/printTestCaseResult.js";import{DecisionCode as a}from"../types/decisionCode.js";import{judgeByStaticAnalysis as o}from"../helpers/judgeByStaticAnalysis.js";import{isSafeSubmissionOutputPath as s}from"../helpers/readOutputFiles.js";import{passesIsolationCheckInDebugMode as c}from"../helpers/runIsolationCheckInDebugMode.js";import{readProblemMarkdownFrontMatter as l}from"../helpers/readProblemMarkdownFrontMatter.js";import{printDebugCwdBanner as u,printDebugExpectationFailureBanner as d,resolveCwds as f}from"../helpers/resolveCwds.js";import p from"node:fs/promises";import m from"node:path";const h={unterminated_quote:'引用符（`"`）が閉じられていません',unexpected_quote:'引用符（`"`）は値全体を囲む位置にしか書けません',text_after_quoted_field:`閉じ引用符の後に区切り文字以外の文字があります`},g={rmsle:{label:`RMSLE`,isHigherBetter:!1,validatePrediction:e=>k(e)??(Number(e)<0?`負の値です`:void 0),compute:(e,t)=>Math.sqrt(A(e.map((e,n)=>(Math.log1p(Number(e))-Math.log1p(Number(t[n])))**2)))},rmse:{label:`RMSE`,isHigherBetter:!1,validatePrediction:k,compute:(e,t)=>Math.sqrt(A(e.map((e,n)=>(Number(e)-Number(t[n]))**2)))},mae:{label:`MAE`,isHigherBetter:!1,validatePrediction:k,compute:(e,t)=>A(e.map((e,n)=>Math.abs(Number(e)-Number(t[n]))))},accuracy:{label:`Accuracy`,isHigherBetter:!0,validatePrediction:e=>e.trim()===``?`空です`:void 0,compute:(e,t)=>A(e.map((e,n)=>+(e.trim()===t[n]?.trim())))}};async function _(t,n){let r=e(process.argv),{cwds:a,isDebugMode:o}=await f(t,r.cwd);if(o&&!await c(t,a,r.params)){process.exitCode=1;return}let s=await C(m.join(t,n.answerFilePath),n);for(let e of a){o&&u(t,e);let r=await v(t,e.cwd,s,n);i({testCaseId:`evaluation`,...r}),o&&!O(e,r)&&(process.exitCode=1,d(t,e))}}async function v(e,t,n,r){let i=await o(t,await l(e));if(i)return i;let c=r.submissionFilePath??`submission.csv`,u;try{let e=m.join(t,c);if(!await s(t,e))throw Error(`unsafe path`);u=await p.readFile(e,`utf8`)}catch{return{decisionCode:a.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,feedbackMarkdown:`\`${c}\`が見つかりません。予測結果のCSVファイルを\`${c}\`という名前で提出してください。`}}let d=S(u,c,n,r);if(typeof d==`string`)return{decisionCode:a.WRONG_ANSWER,feedbackMarkdown:d};let f=[...n.values()],h=r.metric.compute([...n.keys()].map(e=>d.get(e)),f);if(!Number.isFinite(h))return{decisionCode:a.WRONG_ANSWER,feedbackMarkdown:`予測値が極端なため、${r.metric.label}を計算できませんでした。予測値を見直してください。`};let g=r.acceptableScore===void 0||(r.metric.isHigherBetter?h>=r.acceptableScore:h<=r.acceptableScore);return{decisionCode:g?a.ACCEPTED:a.WRONG_ANSWER,score:h,scoreLabel:r.metric.label,feedbackMarkdown:y(h,f.length,g,r)}}function y(e,t,n,r){let{label:i,isHigherBetter:a}=r.metric,o=b(e,r.acceptableScore),s=r.acceptableScore===void 0?``:`\n合格基準: ${i} ${a?`≧`:`≦`} ${x(r.acceptableScore,o)}（${n?`達成`:`未達成`}）`;return`| 指標 | スコア |
 | ---- | ------ |
-| ${i} | ${S(e,o)} |
+| ${i} | ${x(e,o)} |
 
 評価件数: ${t.toLocaleString(`en-US`)}件${s}
-`}function x(e,t){if(t===void 0||e===t)return 6;for(let n=6;n<17;n++)if(S(e,n)!==S(t,n))return n;return 17}function S(e,t){return Number(e.toPrecision(t)).toString()}function C(e,n,i,a){let o;try{o=r(e)}catch(e){return`\`${n}\`をCSVとして読み込めませんでした。${e instanceof t?`${e.lineNumber}行目付近: ${g[e.reason]}`:``}`}let s=T(o[0]?.cells??[],a);if(!s)return`\`${n}\`の1行目には\`${a.idColumn}\`列と\`${a.targetColumn}\`列のヘッダーが必要です。`;let c=new Map,l=[],u=0,d=e=>{u++,l.length<10&&l.push(e)};for(let{cells:e,lineNumber:t}of o.slice(1)){if(E(e))continue;let n=e[s.idIndex]?.trim()??``,r=e[s.targetIndex]??``;if(!i.has(n))d(`${t}行目: \`${a.idColumn}\`が\`${O(n)}\`の行は評価対象ではありません`);else if(c.has(n))d(`${t}行目: \`${a.idColumn}\`が\`${O(n)}\`の行が重複しています`);else{let e=a.metric.validatePrediction?.(r);e&&d(`${t}行目: \`${a.targetColumn}\`の値\`${O(r)}\`は${e}`),c.set(n,r)}}let f=[...i.keys()].filter(e=>!c.has(e));return f.length>0&&(u++,l.unshift(`\`${a.idColumn}\`が${D(f)}の行がありません（不足 ${f.length.toLocaleString(`en-US`)}件）`)),u>0?`\`${n}\`の内容に問題があります。
+`}function b(e,t){if(t===void 0||e===t)return 6;for(let n=6;n<17;n++)if(x(e,n)!==x(t,n))return n;return 17}function x(e,t){return Number(e.toPrecision(t)).toString()}function S(e,n,i,a){let o;try{o=r(e)}catch(e){return`\`${n}\`をCSVとして読み込めませんでした。${e instanceof t?`${e.lineNumber}行目付近: ${h[e.reason]}`:``}`}let s=w(o[0]?.cells??[],a);if(!s)return`\`${n}\`の1行目には\`${a.idColumn}\`列と\`${a.targetColumn}\`列のヘッダーが必要です。`;let c=new Map,l=[],u=0,d=e=>{u++,l.length<10&&l.push(e)};for(let{cells:e,lineNumber:t}of o.slice(1)){if(T(e))continue;let n=e[s.idIndex]?.trim()??``,r=e[s.targetIndex]??``;if(!i.has(n))d(`${t}行目: \`${a.idColumn}\`が\`${D(n)}\`の行は評価対象ではありません`);else if(c.has(n))d(`${t}行目: \`${a.idColumn}\`が\`${D(n)}\`の行が重複しています`);else{let e=a.metric.validatePrediction?.(r);e&&d(`${t}行目: \`${a.targetColumn}\`の値\`${D(r)}\`は${e}`),c.set(n,r)}}let f=[...i.keys()].filter(e=>!c.has(e));return f.length>0&&(u++,l.unshift(`\`${a.idColumn}\`が${E(f)}の行がありません（不足 ${f.length.toLocaleString(`en-US`)}件）`)),u>0?`\`${n}\`の内容に問題があります。
 
 ${l.map(e=>`- ${e}`).join(`
 `)}${u>l.length?`\n- ほか${(u-l.length).toLocaleString(`en-US`)}件`:``}
-`:c}async function w(e,t){let r=n(await m.readFile(e,`utf8`)),i=T(r[0]??[],t);if(!i)throw Error(`answer file must have ${t.idColumn} and ${t.targetColumn} columns: ${e}`);let a=new Map;for(let n of r.slice(1)){if(E(n))continue;let r=n[i.idIndex]?.trim()??``,o=n[i.targetIndex]??``;if(!r)throw Error(`empty ${t.idColumn} in answer file: ${e}`);if(a.has(r))throw Error(`duplicate ${t.idColumn} in answer file: ${r}`);let s=t.metric.validatePrediction?.(o);if(s)throw Error(`invalid ${t.targetColumn} in answer file for ${r}: ${s}`);a.set(r,o)}if(a.size===0)throw Error(`answer file has no rows: ${e}`);return a}function T(e,t){let n=e.map(e=>e.trim()),r=n.indexOf(t.idColumn),i=n.indexOf(t.targetColumn);return r===-1||i===-1?void 0:{idIndex:r,targetIndex:i}}function E(e){return e.every(e=>e.trim()===``)}function D(e){let t=e.slice(0,10).map(e=>`\`${O(e)}\``);return e.length>10?`${t.join(`, `)}, ...`:t.join(`, `)}function O(e){let t=e.replaceAll(/[\p{Cc}]/gu,` `).replaceAll("`",`'`);return t.trim()===``?`（空）`:t.length>50?`${t.slice(0,50)}…`:t}function k(e,t){return t.decisionCode===a.ACCEPTED==(e.expectedResult===`accepted`)}function A(e){return Number.isFinite(Number(e))&&e.trim()!==``?void 0:`数値ではありません`}function j(e){return e.reduce((e,t)=>e+t,0)/e.length}export{v as evaluationJudgePreset,_ as evaluationMetrics};
+`:c}async function C(e,t){let r=n(await p.readFile(e,`utf8`)),i=w(r[0]??[],t);if(!i)throw Error(`answer file must have ${t.idColumn} and ${t.targetColumn} columns: ${e}`);let a=new Map;for(let n of r.slice(1)){if(T(n))continue;let r=n[i.idIndex]?.trim()??``,o=n[i.targetIndex]??``;if(!r)throw Error(`empty ${t.idColumn} in answer file: ${e}`);if(a.has(r))throw Error(`duplicate ${t.idColumn} in answer file: ${r}`);let s=t.metric.validatePrediction?.(o);if(s)throw Error(`invalid ${t.targetColumn} in answer file for ${r}: ${s}`);a.set(r,o)}if(a.size===0)throw Error(`answer file has no rows: ${e}`);return a}function w(e,t){let n=e.map(e=>e.trim()),r=n.indexOf(t.idColumn),i=n.indexOf(t.targetColumn);return r===-1||i===-1?void 0:{idIndex:r,targetIndex:i}}function T(e){return e.every(e=>e.trim()===``)}function E(e){let t=e.slice(0,10).map(e=>`\`${D(e)}\``);return e.length>10?`${t.join(`, `)}, ...`:t.join(`, `)}function D(e){let t=e.replaceAll(/[\p{Cc}]/gu,` `).replaceAll("`",`'`);return t.trim()===``?`（空）`:t.length>50?`${t.slice(0,50)}…`:t}function O(e,t){return t.decisionCode===a.ACCEPTED==(e.expectedResult===`accepted`)}function k(e){return Number.isFinite(Number(e))&&e.trim()!==``?void 0:`数値ではありません`}function A(e){return e.reduce((e,t)=>e+t,0)/e.length}export{_ as evaluationJudgePreset,g as evaluationMetrics};
 //# sourceMappingURL=evaluation.js.map

package/dist/presets/evaluation.js.map

@@ -1 +1 @@
-{"version":3,"file":"evaluation.js","names":[],"sources":["../../src/presets/evaluation.ts"],"sourcesContent":["import fs from 'node:fs/promises';\nimport path from 'node:path';\n\nimport { checkProblemDirIsolation } from '../helpers/checkProblemDirIsolation.js';\nimport { judgeByStaticAnalysis } from '../helpers/judgeByStaticAnalysis.js';\nimport { parseArgs } from '../helpers/parseArgs.js';\nimport { CsvParseError, type CsvParseErrorReason, parseCsv, parseCsvRecords } from '../helpers/parseCsv.js';\nimport { printDebugBanner } from '../helpers/printDebugBanner.js';\nimport { printTestCaseResult } from '../helpers/printTestCaseResult.js';\nimport { isSafeSubmissionOutputPath } from '../helpers/readOutputFiles.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport {\n  printDebugCwdBanner,\n  printDebugExpectationFailureBanner,\n  resolveCwds,\n  type ResolvedCwd,\n} from '../helpers/resolveCwds.js';\nimport { DecisionCode } from '../types/decisionCode.js';\nimport type { TestCaseResult } from '../types/testCaseResult.js';\n\nconst DEFAULT_SUBMISSION_FILE_PATH = 'submission.csv';\nconst TEST_CASE_ID = 'evaluation';\nconst MAX_LISTED_IDS = 10;\n// Learner-controlled cells are echoed into the feedback, so clip them to keep the result line small.\nconst MAX_ECHOED_CELL_LENGTH = 50;\nconst CSV_PARSE_ERROR_MESSAGES: Record<CsvParseErrorReason, string> = {\n  unterminated_quote: '引用符（`\"`）が閉じられていません',\n  unexpected_quote: '引用符（`\"`）は値全体を囲む位置にしか書けません',\n  text_after_quoted_field: '閉じ引用符の後に区切り文字以外の文字があります',\n};\n\nexport interface EvaluationMetric {\n  /** Label shown with the score, e.g. `RMSLE`. */\n  label: string;\n  isHigherBetter: boolean;\n  /**\n   * Returns an error message when a predicted value cannot be scored (e.g. it is not a number).\n   * The message is appended to the row's line number and the offending value, so phrase it as a\n   * predicate such as `数値ではありません`. It also validates the ground truth when it is loaded.\n   */\n  validatePrediction?: (prediction: string) => string | undefined;\n  /** Computes the score from predictions and answers aligned by id. Both are raw CSV cell strings. */\n  compute: (predictions: readonly string[], answers: readonly string[]) => number;\n}\n\nexport interface EvaluationJudgePresetOptions {\n  /** CSV file holding the hidden ground truth, relative to the problem directory. */\n  answerFilePath: string;\n  /** CSV file the submission must contain. Defaults to `submission.csv`. */\n  submissionFilePath?: string;\n  /** Column that identifies each row in both CSV files. */\n  idColumn: string;\n  /** Column holding the answer in the ground truth and the prediction in the submission. */\n  targetColumn: string;\n  metric: EvaluationMetric;\n  /**\n   * Score required for acceptance: the submission is accepted when its score is at least this value\n   * for a higher-is-better metric and at most this value otherwise. Every valid submission is\n   * accepted when omitted.\n   */\n  acceptableScore?: number;\n}\n\nexport const evaluationMetrics = {\n  rmsle: {\n    label: 'RMSLE',\n    isHigherBetter: false,\n    validatePrediction: (prediction: string) =>\n      validateFiniteNumber(prediction) ?? (Number(prediction) < 0 ? '負の値です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(\n        mean(\n          predictions.map(\n            (prediction, index) => (Math.log1p(Number(prediction)) - Math.log1p(Number(answers[index]))) ** 2\n          )\n        )\n      ),\n  },\n  rmse: {\n    label: 'RMSE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(mean(predictions.map((prediction, index) => (Number(prediction) - Number(answers[index])) ** 2))),\n  },\n  mae: {\n    label: 'MAE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => Math.abs(Number(prediction) - Number(answers[index])))),\n  },\n  accuracy: {\n    label: 'Accuracy',\n    isHigherBetter: true,\n    validatePrediction: (prediction: string) => (prediction.trim() === '' ? '空です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => (prediction.trim() === answers[index]?.trim() ? 1 : 0))),\n  },\n} as const satisfies Record<string, EvaluationMetric>;\n\n/**\n * A preset for Kaggle-style model evaluation problems: the submission is a CSV of predictions for a\n * hidden test set, and the judge reports the metric value as `score` instead of comparing outputs.\n * The problem directory keeps the ground truth CSV, so it is never shipped to learners.\n *\n * @example\n * Create `judge.ts`:\n * ```ts\n * import { evaluationJudgePreset, evaluationMetrics } from '@exercode/problem-utils/presets/evaluation';\n *\n * await evaluationJudgePreset(import.meta.dirname, {\n *   answerFilePath: 'evaluation/answer.csv',\n *   idColumn: 'Id',\n *   targetColumn: 'TradePrice',\n *   metric: evaluationMetrics.rmsle,\n *   acceptableScore: 0.5,\n * });\n * ```\n *\n * Run `bun judge.ts model_answers/default` to judge one submission, or `bun judge.ts` to judge every\n * `<problemDir>/model_answers/*` and `<problemDir>/model_answers.fails/*` directory for debugging.\n */\nexport async function evaluationJudgePreset(problemDir: string, options: EvaluationJudgePresetOptions): Promise<void> {\n  const args = parseArgs(process.argv);\n  const { cwds, isDebugMode } = await resolveCwds(problemDir, args.cwd);\n\n  if (isDebugMode) {\n    const acceptedCwd = cwds.find((cwd) => cwd.expectedResult === 'accepted');\n    if (acceptedCwd) {\n      const isolationCheckResult = await checkProblemDirIsolation(problemDir, acceptedCwd, args.params);\n      if (!isolationCheckResult.passed) {\n        process.exitCode = 1;\n        return;\n      }\n    } else {\n      printDebugBanner([\n        '[DEBUG MODE] isolated problem directory check skipped',\n        '',\n        'No accepted model answer is available for checking that the copied judge still accepts a valid submission.',\n      ]);\n    }\n  }\n\n  const answer = await readAnswer(path.join(problemDir, options.answerFilePath), options);\n  for (const resolvedCwd of cwds) {\n    if (isDebugMode) printDebugCwdBanner(problemDir, resolvedCwd);\n    const result = await judgeSubmission(problemDir, resolvedCwd.cwd, answer, options);\n    printTestCaseResult({ testCaseId: TEST_CASE_ID, ...result });\n    if (isDebugMode && !matchesExpectedResult(resolvedCwd, result)) {\n      process.exitCode = 1;\n      printDebugExpectationFailureBanner(problemDir, resolvedCwd);\n    }\n  }\n}\n\nasync function judgeSubmission(\n  problemDir: string,\n  cwd: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Promise<Omit<TestCaseResult, 'testCaseId'>> {\n  const staticAnalysisResult = await judgeByStaticAnalysis(cwd, await readProblemMarkdownFrontMatter(problemDir));\n  if (staticAnalysisResult) return staticAnalysisResult;\n\n  const submissionFilePath = options.submissionFilePath ?? DEFAULT_SUBMISSION_FILE_PATH;\n  let submissionText: string;\n  try {\n    const resolvedPath = path.join(cwd, submissionFilePath);\n    // The harness reads the file as the trusted user, so refuse symlinks escaping the submission.\n    if (!(await isSafeSubmissionOutputPath(cwd, resolvedPath))) throw new Error('unsafe path');\n    submissionText = await fs.readFile(resolvedPath, 'utf8');\n  } catch {\n    return {\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      feedbackMarkdown: `\\`${submissionFilePath}\\`が見つかりません。予測結果のCSVファイルを\\`${submissionFilePath}\\`という名前で提出してください。`,\n    };\n  }\n\n  const predictions = readPredictions(submissionText, submissionFilePath, answer, options);\n  if (typeof predictions === 'string') {\n    return { decisionCode: DecisionCode.WRONG_ANSWER, feedbackMarkdown: predictions };\n  }\n\n  const answerValues = [...answer.values()];\n  const score = options.metric.compute(\n    [...answer.keys()].map((id) => predictions.get(id) as string),\n    answerValues\n  );\n  if (!Number.isFinite(score)) {\n    return {\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      feedbackMarkdown: `予測値が極端なため、${options.metric.label}を計算できませんでした。予測値を見直してください。`,\n    };\n  }\n  const isAccepted =\n    options.acceptableScore === undefined ||\n    (options.metric.isHigherBetter ? score >= options.acceptableScore : score <= options.acceptableScore);\n  return {\n    decisionCode: isAccepted ? DecisionCode.ACCEPTED : DecisionCode.WRONG_ANSWER,\n    score,\n    scoreLabel: options.metric.label,\n    feedbackMarkdown: buildScoreMarkdown(score, answerValues.length, isAccepted, options),\n  };\n}\n\nfunction buildScoreMarkdown(\n  score: number,\n  rowCount: number,\n  isAccepted: boolean,\n  options: EvaluationJudgePresetOptions\n): string {\n  const { label, isHigherBetter } = options.metric;\n  const precision = findDistinguishingPrecision(score, options.acceptableScore);\n  const acceptanceLine =\n    options.acceptableScore === undefined\n      ? ''\n      : `\\n合格基準: ${label} ${isHigherBetter ? '≧' : '≦'} ${formatScore(options.acceptableScore, precision)}（${isAccepted ? '達成' : '未達成'}）`;\n  return `| 指標 | スコア |\n| ---- | ------ |\n| ${label} | ${formatScore(score, precision)} |\n\n評価件数: ${rowCount.toLocaleString('en-US')}件${acceptanceLine}\n`;\n}\n\n/**\n * Finds the number of significant digits at which the score and the threshold render differently\n * (or stop changing), so the feedback never shows a pair that contradicts the decision.\n */\nfunction findDistinguishingPrecision(score: number, threshold: number | undefined): number {\n  const minimumPrecision = 6;\n  if (threshold === undefined || score === threshold) return minimumPrecision;\n  for (let precision = minimumPrecision; precision < 17; precision++) {\n    if (formatScore(score, precision) !== formatScore(threshold, precision)) return precision;\n  }\n  return 17;\n}\n\nfunction formatScore(score: number, precision: number): string {\n  return Number(score.toPrecision(precision)).toString();\n}\n\nfunction readPredictions(\n  submissionText: string,\n  submissionFilePath: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Map<string, string> | string {\n  let records: ReturnType<typeof parseCsvRecords>;\n  try {\n    records = parseCsvRecords(submissionText);\n  } catch (error) {\n    const detail =\n      error instanceof CsvParseError ? `${error.lineNumber}行目付近: ${CSV_PARSE_ERROR_MESSAGES[error.reason]}` : '';\n    return `\\`${submissionFilePath}\\`をCSVとして読み込めませんでした。${detail}`;\n  }\n  const columns = findColumns(records[0]?.cells ?? [], options);\n  if (!columns) {\n    return `\\`${submissionFilePath}\\`の1行目には\\`${options.idColumn}\\`列と\\`${options.targetColumn}\\`列のヘッダーが必要です。`;\n  }\n\n  const predictions = new Map<string, string>();\n  const problems: string[] = [];\n  let problemCount = 0;\n  const reportProblem = (problem: string): void => {\n    problemCount++;\n    if (problems.length < MAX_LISTED_IDS) problems.push(problem);\n  };\n  for (const { cells: row, lineNumber } of records.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const prediction = row[columns.targetIndex] ?? '';\n    if (!answer.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行は評価対象ではありません`);\n    } else if (predictions.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行が重複しています`);\n    } else {\n      const validationError = options.metric.validatePrediction?.(prediction);\n      if (validationError) {\n        reportProblem(`${lineNumber}行目: \\`${options.targetColumn}\\`の値\\`${clip(prediction)}\\`は${validationError}`);\n      }\n      predictions.set(id, prediction);\n    }\n  }\n  const missingIds = [...answer.keys()].filter((id) => !predictions.has(id));\n  if (missingIds.length > 0) {\n    // Missing rows are the most actionable problem, so they must survive the listing cap.\n    problemCount++;\n    problems.unshift(\n      `\\`${options.idColumn}\\`が${listIds(missingIds)}の行がありません（不足 ${missingIds.length.toLocaleString('en-US')}件）`\n    );\n  }\n  if (problemCount > 0) {\n    return `\\`${submissionFilePath}\\`の内容に問題があります。\n\n${problems.map((problem) => `- ${problem}`).join('\\n')}${problemCount > problems.length ? `\\n- ほか${(problemCount - problems.length).toLocaleString('en-US')}件` : ''}\n`;\n  }\n  return predictions;\n}\n\nasync function readAnswer(answerFilePath: string, options: EvaluationJudgePresetOptions): Promise<Map<string, string>> {\n  const rows = parseCsv(await fs.readFile(answerFilePath, 'utf8'));\n  const columns = findColumns(rows[0] ?? [], options);\n  if (!columns) {\n    throw new Error(`answer file must have ${options.idColumn} and ${options.targetColumn} columns: ${answerFilePath}`);\n  }\n  const answer = new Map<string, string>();\n  for (const row of rows.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const value = row[columns.targetIndex] ?? '';\n    if (!id) throw new Error(`empty ${options.idColumn} in answer file: ${answerFilePath}`);\n    if (answer.has(id)) throw new Error(`duplicate ${options.idColumn} in answer file: ${id}`);\n    const validationError = options.metric.validatePrediction?.(value);\n    if (validationError)\n      throw new Error(`invalid ${options.targetColumn} in answer file for ${id}: ${validationError}`);\n    answer.set(id, value);\n  }\n  if (answer.size === 0) throw new Error(`answer file has no rows: ${answerFilePath}`);\n  return answer;\n}\n\nfunction findColumns(\n  header: readonly string[],\n  options: EvaluationJudgePresetOptions\n): { idIndex: number; targetIndex: number } | undefined {\n  const names = header.map((name) => name.trim());\n  const idIndex = names.indexOf(options.idColumn);\n  const targetIndex = names.indexOf(options.targetColumn);\n  return idIndex === -1 || targetIndex === -1 ? undefined : { idIndex, targetIndex };\n}\n\nfunction isBlankRow(row: readonly string[]): boolean {\n  return row.every((cell) => cell.trim() === '');\n}\n\nfunction listIds(ids: readonly string[]): string {\n  const listed = ids.slice(0, MAX_LISTED_IDS).map((id) => `\\`${clip(id)}\\``);\n  return ids.length > MAX_LISTED_IDS ? `${listed.join(', ')}, ...` : listed.join(', ');\n}\n\nfunction clip(value: string): string {\n  // Control characters and backticks would break the markdown list item and inline code the value is echoed in.\n  const printable = value.replaceAll(/[\\p{Cc}]/gu, ' ').replaceAll('`', \"'\");\n  if (printable.trim() === '') return '（空）';\n  return printable.length > MAX_ECHOED_CELL_LENGTH ? `${printable.slice(0, MAX_ECHOED_CELL_LENGTH)}…` : printable;\n}\n\nfunction matchesExpectedResult(resolvedCwd: ResolvedCwd, result: Pick<TestCaseResult, 'decisionCode'>): boolean {\n  return (result.decisionCode === DecisionCode.ACCEPTED) === (resolvedCwd.expectedResult === 'accepted');\n}\n\nfunction validateFiniteNumber(prediction: string): string | undefined {\n  return Number.isFinite(Number(prediction)) && prediction.trim() !== '' ? undefined : '数値ではありません';\n}\n\nfunction mean(values: readonly number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length;\n}\n"],"mappings":"y0BAoBA,MAKM,EAAgE,CACpE,mBAAoB,qBACpB,iBAAkB,4BAClB,wBAAyB,yBAC3B,EAkCa,EAAoB,CAC/B,MAAO,CACL,MAAO,QACP,eAAgB,GAChB,mBAAqB,GACnB,EAAqB,CAAU,IAAM,OAAO,CAAU,EAAI,EAAI,QAAU,IAAA,IAC1E,SAAU,EAAgC,IACxC,KAAK,KACH,EACE,EAAY,KACT,EAAY,KAAW,KAAK,MAAM,OAAO,CAAU,CAAC,EAAI,KAAK,MAAM,OAAO,EAAQ,EAAM,CAAC,IAAM,CAClG,CACF,CACF,CACJ,EACA,KAAM,CACJ,MAAO,OACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,KAAK,KAAK,EAAK,EAAY,KAAK,EAAY,KAAW,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,IAAM,CAAC,CAAC,CAAC,CAC9G,EACA,IAAK,CACH,MAAO,MACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAU,KAAK,IAAI,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,CAAC,CAAC,CAAC,CACtG,EACA,SAAU,CACR,MAAO,WACP,eAAgB,GAChB,mBAAqB,GAAwB,EAAW,KAAK,IAAM,GAAK,MAAQ,IAAA,GAChF,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAW,IAAW,KAAK,IAAM,EAAQ,EAAM,EAAE,KAAK,EAAU,CAAC,CACvG,CACF,EAwBA,eAAsB,EAAsB,EAAoB,EAAsD,CACpH,IAAM,EAAO,EAAU,QAAQ,IAAI,EAC7B,CAAE,OAAM,eAAgB,MAAM,EAAY,EAAY,EAAK,GAAG,EAEpE,GAAI,EAAa,CACf,IAAM,EAAc,EAAK,KAAM,GAAQ,EAAI,iBAAmB,UAAU,EACxE,GAAI,EAEE,IAAA,EAAC,MAD8B,EAAyB,EAAY,EAAa,EAAK,MAAM,EAAA,CACtE,OAAQ,CAChC,QAAQ,SAAW,EACnB,MACF,OAEA,EAAiB,CACf,wDACA,GACA,4GACF,CAAC,CAEL,CAEA,IAAM,EAAS,MAAM,EAAW,EAAK,KAAK,EAAY,EAAQ,cAAc,EAAG,CAAO,EACtF,IAAK,IAAM,KAAe,EAAM,CAC1B,GAAa,EAAoB,EAAY,CAAW,EAC5D,IAAM,EAAS,MAAM,EAAgB,EAAY,EAAY,IAAK,EAAQ,CAAO,EACjF,EAAoB,CAAE,WAAY,aAAc,GAAG,CAAO,CAAC,EACvD,GAAe,CAAC,EAAsB,EAAa,CAAM,IAC3D,QAAQ,SAAW,EACnB,EAAmC,EAAY,CAAW,EAE9D,CACF,CAEA,eAAe,EACb,EACA,EACA,EACA,EAC6C,CAC7C,IAAM,EAAuB,MAAM,EAAsB,EAAK,MAAM,EAA+B,CAAU,CAAC,EAC9G,GAAI,EAAsB,OAAO,EAEjC,IAAM,EAAqB,EAAQ,oBAAsB,iBACrD,EACJ,GAAI,CACF,IAAM,EAAe,EAAK,KAAK,EAAK,CAAkB,EAEtD,GAAI,CAAE,MAAM,EAA2B,EAAK,CAAY,EAAI,MAAU,MAAM,aAAa,EACzF,EAAiB,MAAM,EAAG,SAAS,EAAc,MAAM,CACzD,MAAQ,CACN,MAAO,CACL,aAAc,EAAa,uCAC3B,iBAAkB,KAAK,EAAmB,4BAA4B,EAAmB,kBAC3F,CACF,CAEA,IAAM,EAAc,EAAgB,EAAgB,EAAoB,EAAQ,CAAO,EACvF,GAAI,OAAO,GAAgB,SACzB,MAAO,CAAE,aAAc,EAAa,aAAc,iBAAkB,CAAY,EAGlF,IAAM,EAAe,CAAC,GAAG,EAAO,OAAO,CAAC,EAClC,EAAQ,EAAQ,OAAO,QAC3B,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,IAAK,GAAO,EAAY,IAAI,CAAE,CAAW,EAC5D,CACF,EACA,GAAI,CAAC,OAAO,SAAS,CAAK,EACxB,MAAO,CACL,aAAc,EAAa,aAC3B,iBAAkB,aAAa,EAAQ,OAAO,MAAM,0BACtD,EAEF,IAAM,EACJ,EAAQ,kBAAoB,IAAA,KAC3B,EAAQ,OAAO,eAAiB,GAAS,EAAQ,gBAAkB,GAAS,EAAQ,iBACvF,MAAO,CACL,aAAc,EAAa,EAAa,SAAW,EAAa,aAChE,QACA,WAAY,EAAQ,OAAO,MAC3B,iBAAkB,EAAmB,EAAO,EAAa,OAAQ,EAAY,CAAO,CACtF,CACF,CAEA,SAAS,EACP,EACA,EACA,EACA,EACQ,CACR,GAAM,CAAE,QAAO,kBAAmB,EAAQ,OACpC,EAAY,EAA4B,EAAO,EAAQ,eAAe,EACtE,EACJ,EAAQ,kBAAoB,IAAA,GACxB,GACA,WAAW,EAAM,GAAG,EAAiB,IAAM,IAAI,GAAG,EAAY,EAAQ,gBAAiB,CAAS,EAAE,GAAG,EAAa,KAAO,MAAM,GACrI,MAAO;;IAEL,EAAM,KAAK,EAAY,EAAO,CAAS,EAAE;;QAErC,EAAS,eAAe,OAAO,EAAE,GAAG,EAAe;CAE3D,CAMA,SAAS,EAA4B,EAAe,EAAuC,CAEzF,GAAI,IAAc,IAAA,IAAa,IAAU,EAAW,MAAO,GAC3D,IAAK,IAAI,EAAY,EAAkB,EAAY,GAAI,IACrD,GAAI,EAAY,EAAO,CAAS,IAAM,EAAY,EAAW,CAAS,EAAG,OAAO,EAElF,MAAO,GACT,CAEA,SAAS,EAAY,EAAe,EAA2B,CAC7D,OAAO,OAAO,EAAM,YAAY,CAAS,CAAC,CAAC,CAAC,SAAS,CACvD,CAEA,SAAS,EACP,EACA,EACA,EACA,EAC8B,CAC9B,IAAI,EACJ,GAAI,CACF,EAAU,EAAgB,CAAc,CAC1C,OAAS,EAAO,CAGd,MAAO,KAAK,EAAmB,sBAD7B,aAAiB,EAAgB,GAAG,EAAM,WAAW,QAAQ,EAAyB,EAAM,UAAY,IAE5G,CACA,IAAM,EAAU,EAAY,EAAQ,EAAE,EAAE,OAAS,CAAC,EAAG,CAAO,EAC5D,GAAI,CAAC,EACH,MAAO,KAAK,EAAmB,YAAY,EAAQ,SAAS,QAAQ,EAAQ,aAAa,gBAG3F,IAAM,EAAc,IAAI,IAClB,EAAqB,CAAC,EACxB,EAAe,EACb,EAAiB,GAA0B,CAC/C,IACI,EAAS,OAAS,IAAgB,EAAS,KAAK,CAAO,CAC7D,EACA,IAAK,GAAM,CAAE,MAAO,EAAK,gBAAgB,EAAQ,MAAM,CAAC,EAAG,CACzD,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAa,EAAI,EAAQ,cAAgB,GAC/C,GAAI,CAAC,EAAO,IAAI,CAAE,EAChB,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,iBAAiB,OACjF,GAAI,EAAY,IAAI,CAAE,EAC3B,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,aAAa,MAC7E,CACL,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAU,EAClE,GACF,EAAc,GAAG,EAAW,QAAQ,EAAQ,aAAa,QAAQ,EAAK,CAAU,EAAE,KAAK,GAAiB,EAE1G,EAAY,IAAI,EAAI,CAAU,CAChC,CACF,CACA,IAAM,EAAa,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,OAAQ,GAAO,CAAC,EAAY,IAAI,CAAE,CAAC,EAczE,OAbI,EAAW,OAAS,IAEtB,IACA,EAAS,QACP,KAAK,EAAQ,SAAS,KAAK,EAAQ,CAAU,EAAE,cAAc,EAAW,OAAO,eAAe,OAAO,EAAE,GACzG,GAEE,EAAe,EACV,KAAK,EAAmB;;EAEjC,EAAS,IAAK,GAAY,KAAK,GAAS,CAAC,CAAC,KAAK;CAAI,IAAI,EAAe,EAAS,OAAS,UAAU,EAAe,EAAS,OAAA,CAAQ,eAAe,OAAO,EAAE,GAAK,GAAG;EAG3J,CACT,CAEA,eAAe,EAAW,EAAwB,EAAqE,CACrH,IAAM,EAAO,EAAS,MAAM,EAAG,SAAS,EAAgB,MAAM,CAAC,EACzD,EAAU,EAAY,EAAK,IAAM,CAAC,EAAG,CAAO,EAClD,GAAI,CAAC,EACH,MAAU,MAAM,yBAAyB,EAAQ,SAAS,OAAO,EAAQ,aAAa,YAAY,GAAgB,EAEpH,IAAM,EAAS,IAAI,IACnB,IAAK,IAAM,KAAO,EAAK,MAAM,CAAC,EAAG,CAC/B,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAQ,EAAI,EAAQ,cAAgB,GAC1C,GAAI,CAAC,EAAI,MAAU,MAAM,SAAS,EAAQ,SAAS,mBAAmB,GAAgB,EACtF,GAAI,EAAO,IAAI,CAAE,EAAG,MAAU,MAAM,aAAa,EAAQ,SAAS,mBAAmB,GAAI,EACzF,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAK,EACjE,GAAI,EACF,MAAU,MAAM,WAAW,EAAQ,aAAa,sBAAsB,EAAG,IAAI,GAAiB,EAChG,EAAO,IAAI,EAAI,CAAK,CACtB,CACA,GAAI,EAAO,OAAS,EAAG,MAAU,MAAM,4BAA4B,GAAgB,EACnF,OAAO,CACT,CAEA,SAAS,EACP,EACA,EACsD,CACtD,IAAM,EAAQ,EAAO,IAAK,GAAS,EAAK,KAAK,CAAC,EACxC,EAAU,EAAM,QAAQ,EAAQ,QAAQ,EACxC,EAAc,EAAM,QAAQ,EAAQ,YAAY,EACtD,OAAO,IAAY,IAAM,IAAgB,GAAK,IAAA,GAAY,CAAE,UAAS,aAAY,CACnF,CAEA,SAAS,EAAW,EAAiC,CACnD,OAAO,EAAI,MAAO,GAAS,EAAK,KAAK,IAAM,EAAE,CAC/C,CAEA,SAAS,EAAQ,EAAgC,CAC/C,IAAM,EAAS,EAAI,MAAM,EAAG,EAAc,CAAC,CAAC,IAAK,GAAO,KAAK,EAAK,CAAE,EAAE,GAAG,EACzE,OAAO,EAAI,OAAS,GAAiB,GAAG,EAAO,KAAK,IAAI,EAAE,OAAS,EAAO,KAAK,IAAI,CACrF,CAEA,SAAS,EAAK,EAAuB,CAEnC,IAAM,EAAY,EAAM,WAAW,aAAc,GAAG,CAAC,CAAC,WAAW,IAAK,GAAG,EAEzE,OADI,EAAU,KAAK,IAAM,GAAW,MAC7B,EAAU,OAAS,GAAyB,GAAG,EAAU,MAAM,EAAG,EAAsB,EAAE,GAAK,CACxG,CAEA,SAAS,EAAsB,EAA0B,EAAuD,CAC9G,OAAQ,EAAO,eAAiB,EAAa,WAAe,EAAY,iBAAmB,WAC7F,CAEA,SAAS,EAAqB,EAAwC,CACpE,OAAO,OAAO,SAAS,OAAO,CAAU,CAAC,GAAK,EAAW,KAAK,IAAM,GAAK,IAAA,GAAY,WACvF,CAEA,SAAS,EAAK,EAAmC,CAC/C,OAAO,EAAO,QAAQ,EAAK,IAAU,EAAM,EAAO,CAAC,EAAI,EAAO,MAChE"}
+{"version":3,"file":"evaluation.js","names":[],"sources":["../../src/presets/evaluation.ts"],"sourcesContent":["import fs from 'node:fs/promises';\nimport path from 'node:path';\n\nimport { judgeByStaticAnalysis } from '../helpers/judgeByStaticAnalysis.js';\nimport { parseArgs } from '../helpers/parseArgs.js';\nimport { CsvParseError, type CsvParseErrorReason, parseCsv, parseCsvRecords } from '../helpers/parseCsv.js';\nimport { printTestCaseResult } from '../helpers/printTestCaseResult.js';\nimport { isSafeSubmissionOutputPath } from '../helpers/readOutputFiles.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport { passesIsolationCheckInDebugMode } from '../helpers/runIsolationCheckInDebugMode.js';\nimport {\n  printDebugCwdBanner,\n  printDebugExpectationFailureBanner,\n  resolveCwds,\n  type ResolvedCwd,\n} from '../helpers/resolveCwds.js';\nimport { DecisionCode } from '../types/decisionCode.js';\nimport type { TestCaseResult } from '../types/testCaseResult.js';\n\nconst DEFAULT_SUBMISSION_FILE_PATH = 'submission.csv';\nconst TEST_CASE_ID = 'evaluation';\nconst MAX_LISTED_IDS = 10;\n// Learner-controlled cells are echoed into the feedback, so clip them to keep the result line small.\nconst MAX_ECHOED_CELL_LENGTH = 50;\nconst CSV_PARSE_ERROR_MESSAGES: Record<CsvParseErrorReason, string> = {\n  unterminated_quote: '引用符（`\"`）が閉じられていません',\n  unexpected_quote: '引用符（`\"`）は値全体を囲む位置にしか書けません',\n  text_after_quoted_field: '閉じ引用符の後に区切り文字以外の文字があります',\n};\n\nexport interface EvaluationMetric {\n  /** Label shown with the score, e.g. `RMSLE`. */\n  label: string;\n  isHigherBetter: boolean;\n  /**\n   * Returns an error message when a predicted value cannot be scored (e.g. it is not a number).\n   * The message is appended to the row's line number and the offending value, so phrase it as a\n   * predicate such as `数値ではありません`. It also validates the ground truth when it is loaded.\n   */\n  validatePrediction?: (prediction: string) => string | undefined;\n  /** Computes the score from predictions and answers aligned by id. Both are raw CSV cell strings. */\n  compute: (predictions: readonly string[], answers: readonly string[]) => number;\n}\n\nexport interface EvaluationJudgePresetOptions {\n  /** CSV file holding the hidden ground truth, relative to the problem directory. */\n  answerFilePath: string;\n  /** CSV file the submission must contain. Defaults to `submission.csv`. */\n  submissionFilePath?: string;\n  /** Column that identifies each row in both CSV files. */\n  idColumn: string;\n  /** Column holding the answer in the ground truth and the prediction in the submission. */\n  targetColumn: string;\n  metric: EvaluationMetric;\n  /**\n   * Score required for acceptance: the submission is accepted when its score is at least this value\n   * for a higher-is-better metric and at most this value otherwise. Every valid submission is\n   * accepted when omitted.\n   */\n  acceptableScore?: number;\n}\n\nexport const evaluationMetrics = {\n  rmsle: {\n    label: 'RMSLE',\n    isHigherBetter: false,\n    validatePrediction: (prediction: string) =>\n      validateFiniteNumber(prediction) ?? (Number(prediction) < 0 ? '負の値です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(\n        mean(\n          predictions.map(\n            (prediction, index) => (Math.log1p(Number(prediction)) - Math.log1p(Number(answers[index]))) ** 2\n          )\n        )\n      ),\n  },\n  rmse: {\n    label: 'RMSE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(mean(predictions.map((prediction, index) => (Number(prediction) - Number(answers[index])) ** 2))),\n  },\n  mae: {\n    label: 'MAE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => Math.abs(Number(prediction) - Number(answers[index])))),\n  },\n  accuracy: {\n    label: 'Accuracy',\n    isHigherBetter: true,\n    validatePrediction: (prediction: string) => (prediction.trim() === '' ? '空です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => (prediction.trim() === answers[index]?.trim() ? 1 : 0))),\n  },\n} as const satisfies Record<string, EvaluationMetric>;\n\n/**\n * A preset for Kaggle-style model evaluation problems: the submission is a CSV of predictions for a\n * hidden test set, and the judge reports the metric value as `score` instead of comparing outputs.\n * The problem directory keeps the ground truth CSV, so it is never shipped to learners.\n *\n * @example\n * Create `judge.ts`:\n * ```ts\n * import { evaluationJudgePreset, evaluationMetrics } from '@exercode/problem-utils/presets/evaluation';\n *\n * await evaluationJudgePreset(import.meta.dirname, {\n *   answerFilePath: 'evaluation/answer.csv',\n *   idColumn: 'Id',\n *   targetColumn: 'TradePrice',\n *   metric: evaluationMetrics.rmsle,\n *   acceptableScore: 0.5,\n * });\n * ```\n *\n * Run `bun judge.ts model_answers/default` to judge one submission, or `bun judge.ts` to judge every\n * `<problemDir>/model_answers/*` and `<problemDir>/model_answers.fails/*` directory for debugging.\n */\nexport async function evaluationJudgePreset(problemDir: string, options: EvaluationJudgePresetOptions): Promise<void> {\n  const args = parseArgs(process.argv);\n  const { cwds, isDebugMode } = await resolveCwds(problemDir, args.cwd);\n\n  if (isDebugMode && !(await passesIsolationCheckInDebugMode(problemDir, cwds, args.params))) {\n    process.exitCode = 1;\n    return;\n  }\n\n  const answer = await readAnswer(path.join(problemDir, options.answerFilePath), options);\n  for (const resolvedCwd of cwds) {\n    if (isDebugMode) printDebugCwdBanner(problemDir, resolvedCwd);\n    const result = await judgeSubmission(problemDir, resolvedCwd.cwd, answer, options);\n    printTestCaseResult({ testCaseId: TEST_CASE_ID, ...result });\n    if (isDebugMode && !matchesExpectedResult(resolvedCwd, result)) {\n      process.exitCode = 1;\n      printDebugExpectationFailureBanner(problemDir, resolvedCwd);\n    }\n  }\n}\n\nasync function judgeSubmission(\n  problemDir: string,\n  cwd: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Promise<Omit<TestCaseResult, 'testCaseId'>> {\n  const staticAnalysisResult = await judgeByStaticAnalysis(cwd, await readProblemMarkdownFrontMatter(problemDir));\n  if (staticAnalysisResult) return staticAnalysisResult;\n\n  const submissionFilePath = options.submissionFilePath ?? DEFAULT_SUBMISSION_FILE_PATH;\n  let submissionText: string;\n  try {\n    const resolvedPath = path.join(cwd, submissionFilePath);\n    // The harness reads the file as the trusted user, so refuse symlinks escaping the submission.\n    if (!(await isSafeSubmissionOutputPath(cwd, resolvedPath))) throw new Error('unsafe path');\n    submissionText = await fs.readFile(resolvedPath, 'utf8');\n  } catch {\n    return {\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      feedbackMarkdown: `\\`${submissionFilePath}\\`が見つかりません。予測結果のCSVファイルを\\`${submissionFilePath}\\`という名前で提出してください。`,\n    };\n  }\n\n  const predictions = readPredictions(submissionText, submissionFilePath, answer, options);\n  if (typeof predictions === 'string') {\n    return { decisionCode: DecisionCode.WRONG_ANSWER, feedbackMarkdown: predictions };\n  }\n\n  const answerValues = [...answer.values()];\n  const score = options.metric.compute(\n    [...answer.keys()].map((id) => predictions.get(id) as string),\n    answerValues\n  );\n  if (!Number.isFinite(score)) {\n    return {\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      feedbackMarkdown: `予測値が極端なため、${options.metric.label}を計算できませんでした。予測値を見直してください。`,\n    };\n  }\n  const isAccepted =\n    options.acceptableScore === undefined ||\n    (options.metric.isHigherBetter ? score >= options.acceptableScore : score <= options.acceptableScore);\n  return {\n    decisionCode: isAccepted ? DecisionCode.ACCEPTED : DecisionCode.WRONG_ANSWER,\n    score,\n    scoreLabel: options.metric.label,\n    feedbackMarkdown: buildScoreMarkdown(score, answerValues.length, isAccepted, options),\n  };\n}\n\nfunction buildScoreMarkdown(\n  score: number,\n  rowCount: number,\n  isAccepted: boolean,\n  options: EvaluationJudgePresetOptions\n): string {\n  const { label, isHigherBetter } = options.metric;\n  const precision = findDistinguishingPrecision(score, options.acceptableScore);\n  const acceptanceLine =\n    options.acceptableScore === undefined\n      ? ''\n      : `\\n合格基準: ${label} ${isHigherBetter ? '≧' : '≦'} ${formatScore(options.acceptableScore, precision)}（${isAccepted ? '達成' : '未達成'}）`;\n  return `| 指標 | スコア |\n| ---- | ------ |\n| ${label} | ${formatScore(score, precision)} |\n\n評価件数: ${rowCount.toLocaleString('en-US')}件${acceptanceLine}\n`;\n}\n\n/**\n * Finds the number of significant digits at which the score and the threshold render differently\n * (or stop changing), so the feedback never shows a pair that contradicts the decision.\n */\nfunction findDistinguishingPrecision(score: number, threshold: number | undefined): number {\n  const minimumPrecision = 6;\n  if (threshold === undefined || score === threshold) return minimumPrecision;\n  for (let precision = minimumPrecision; precision < 17; precision++) {\n    if (formatScore(score, precision) !== formatScore(threshold, precision)) return precision;\n  }\n  return 17;\n}\n\nfunction formatScore(score: number, precision: number): string {\n  return Number(score.toPrecision(precision)).toString();\n}\n\nfunction readPredictions(\n  submissionText: string,\n  submissionFilePath: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Map<string, string> | string {\n  let records: ReturnType<typeof parseCsvRecords>;\n  try {\n    records = parseCsvRecords(submissionText);\n  } catch (error) {\n    const detail =\n      error instanceof CsvParseError ? `${error.lineNumber}行目付近: ${CSV_PARSE_ERROR_MESSAGES[error.reason]}` : '';\n    return `\\`${submissionFilePath}\\`をCSVとして読み込めませんでした。${detail}`;\n  }\n  const columns = findColumns(records[0]?.cells ?? [], options);\n  if (!columns) {\n    return `\\`${submissionFilePath}\\`の1行目には\\`${options.idColumn}\\`列と\\`${options.targetColumn}\\`列のヘッダーが必要です。`;\n  }\n\n  const predictions = new Map<string, string>();\n  const problems: string[] = [];\n  let problemCount = 0;\n  const reportProblem = (problem: string): void => {\n    problemCount++;\n    if (problems.length < MAX_LISTED_IDS) problems.push(problem);\n  };\n  for (const { cells: row, lineNumber } of records.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const prediction = row[columns.targetIndex] ?? '';\n    if (!answer.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行は評価対象ではありません`);\n    } else if (predictions.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行が重複しています`);\n    } else {\n      const validationError = options.metric.validatePrediction?.(prediction);\n      if (validationError) {\n        reportProblem(`${lineNumber}行目: \\`${options.targetColumn}\\`の値\\`${clip(prediction)}\\`は${validationError}`);\n      }\n      predictions.set(id, prediction);\n    }\n  }\n  const missingIds = [...answer.keys()].filter((id) => !predictions.has(id));\n  if (missingIds.length > 0) {\n    // Missing rows are the most actionable problem, so they must survive the listing cap.\n    problemCount++;\n    problems.unshift(\n      `\\`${options.idColumn}\\`が${listIds(missingIds)}の行がありません（不足 ${missingIds.length.toLocaleString('en-US')}件）`\n    );\n  }\n  if (problemCount > 0) {\n    return `\\`${submissionFilePath}\\`の内容に問題があります。\n\n${problems.map((problem) => `- ${problem}`).join('\\n')}${problemCount > problems.length ? `\\n- ほか${(problemCount - problems.length).toLocaleString('en-US')}件` : ''}\n`;\n  }\n  return predictions;\n}\n\nasync function readAnswer(answerFilePath: string, options: EvaluationJudgePresetOptions): Promise<Map<string, string>> {\n  const rows = parseCsv(await fs.readFile(answerFilePath, 'utf8'));\n  const columns = findColumns(rows[0] ?? [], options);\n  if (!columns) {\n    throw new Error(`answer file must have ${options.idColumn} and ${options.targetColumn} columns: ${answerFilePath}`);\n  }\n  const answer = new Map<string, string>();\n  for (const row of rows.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const value = row[columns.targetIndex] ?? '';\n    if (!id) throw new Error(`empty ${options.idColumn} in answer file: ${answerFilePath}`);\n    if (answer.has(id)) throw new Error(`duplicate ${options.idColumn} in answer file: ${id}`);\n    const validationError = options.metric.validatePrediction?.(value);\n    if (validationError)\n      throw new Error(`invalid ${options.targetColumn} in answer file for ${id}: ${validationError}`);\n    answer.set(id, value);\n  }\n  if (answer.size === 0) throw new Error(`answer file has no rows: ${answerFilePath}`);\n  return answer;\n}\n\nfunction findColumns(\n  header: readonly string[],\n  options: EvaluationJudgePresetOptions\n): { idIndex: number; targetIndex: number } | undefined {\n  const names = header.map((name) => name.trim());\n  const idIndex = names.indexOf(options.idColumn);\n  const targetIndex = names.indexOf(options.targetColumn);\n  return idIndex === -1 || targetIndex === -1 ? undefined : { idIndex, targetIndex };\n}\n\nfunction isBlankRow(row: readonly string[]): boolean {\n  return row.every((cell) => cell.trim() === '');\n}\n\nfunction listIds(ids: readonly string[]): string {\n  const listed = ids.slice(0, MAX_LISTED_IDS).map((id) => `\\`${clip(id)}\\``);\n  return ids.length > MAX_LISTED_IDS ? `${listed.join(', ')}, ...` : listed.join(', ');\n}\n\nfunction clip(value: string): string {\n  // Control characters and backticks would break the markdown list item and inline code the value is echoed in.\n  const printable = value.replaceAll(/[\\p{Cc}]/gu, ' ').replaceAll('`', \"'\");\n  if (printable.trim() === '') return '（空）';\n  return printable.length > MAX_ECHOED_CELL_LENGTH ? `${printable.slice(0, MAX_ECHOED_CELL_LENGTH)}…` : printable;\n}\n\nfunction matchesExpectedResult(resolvedCwd: ResolvedCwd, result: Pick<TestCaseResult, 'decisionCode'>): boolean {\n  return (result.decisionCode === DecisionCode.ACCEPTED) === (resolvedCwd.expectedResult === 'accepted');\n}\n\nfunction validateFiniteNumber(prediction: string): string | undefined {\n  return Number.isFinite(Number(prediction)) && prediction.trim() !== '' ? undefined : '数値ではありません';\n}\n\nfunction mean(values: readonly number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length;\n}\n"],"mappings":"kxBAmBA,MAKM,EAAgE,CACpE,mBAAoB,qBACpB,iBAAkB,4BAClB,wBAAyB,yBAC3B,EAkCa,EAAoB,CAC/B,MAAO,CACL,MAAO,QACP,eAAgB,GAChB,mBAAqB,GACnB,EAAqB,CAAU,IAAM,OAAO,CAAU,EAAI,EAAI,QAAU,IAAA,IAC1E,SAAU,EAAgC,IACxC,KAAK,KACH,EACE,EAAY,KACT,EAAY,KAAW,KAAK,MAAM,OAAO,CAAU,CAAC,EAAI,KAAK,MAAM,OAAO,EAAQ,EAAM,CAAC,IAAM,CAClG,CACF,CACF,CACJ,EACA,KAAM,CACJ,MAAO,OACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,KAAK,KAAK,EAAK,EAAY,KAAK,EAAY,KAAW,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,IAAM,CAAC,CAAC,CAAC,CAC9G,EACA,IAAK,CACH,MAAO,MACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAU,KAAK,IAAI,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,CAAC,CAAC,CAAC,CACtG,EACA,SAAU,CACR,MAAO,WACP,eAAgB,GAChB,mBAAqB,GAAwB,EAAW,KAAK,IAAM,GAAK,MAAQ,IAAA,GAChF,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAW,IAAW,KAAK,IAAM,EAAQ,EAAM,EAAE,KAAK,EAAU,CAAC,CACvG,CACF,EAwBA,eAAsB,EAAsB,EAAoB,EAAsD,CACpH,IAAM,EAAO,EAAU,QAAQ,IAAI,EAC7B,CAAE,OAAM,eAAgB,MAAM,EAAY,EAAY,EAAK,GAAG,EAEpE,GAAI,GAAe,CAAE,MAAM,EAAgC,EAAY,EAAM,EAAK,MAAM,EAAI,CAC1F,QAAQ,SAAW,EACnB,MACF,CAEA,IAAM,EAAS,MAAM,EAAW,EAAK,KAAK,EAAY,EAAQ,cAAc,EAAG,CAAO,EACtF,IAAK,IAAM,KAAe,EAAM,CAC1B,GAAa,EAAoB,EAAY,CAAW,EAC5D,IAAM,EAAS,MAAM,EAAgB,EAAY,EAAY,IAAK,EAAQ,CAAO,EACjF,EAAoB,CAAE,WAAY,aAAc,GAAG,CAAO,CAAC,EACvD,GAAe,CAAC,EAAsB,EAAa,CAAM,IAC3D,QAAQ,SAAW,EACnB,EAAmC,EAAY,CAAW,EAE9D,CACF,CAEA,eAAe,EACb,EACA,EACA,EACA,EAC6C,CAC7C,IAAM,EAAuB,MAAM,EAAsB,EAAK,MAAM,EAA+B,CAAU,CAAC,EAC9G,GAAI,EAAsB,OAAO,EAEjC,IAAM,EAAqB,EAAQ,oBAAsB,iBACrD,EACJ,GAAI,CACF,IAAM,EAAe,EAAK,KAAK,EAAK,CAAkB,EAEtD,GAAI,CAAE,MAAM,EAA2B,EAAK,CAAY,EAAI,MAAU,MAAM,aAAa,EACzF,EAAiB,MAAM,EAAG,SAAS,EAAc,MAAM,CACzD,MAAQ,CACN,MAAO,CACL,aAAc,EAAa,uCAC3B,iBAAkB,KAAK,EAAmB,4BAA4B,EAAmB,kBAC3F,CACF,CAEA,IAAM,EAAc,EAAgB,EAAgB,EAAoB,EAAQ,CAAO,EACvF,GAAI,OAAO,GAAgB,SACzB,MAAO,CAAE,aAAc,EAAa,aAAc,iBAAkB,CAAY,EAGlF,IAAM,EAAe,CAAC,GAAG,EAAO,OAAO,CAAC,EAClC,EAAQ,EAAQ,OAAO,QAC3B,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,IAAK,GAAO,EAAY,IAAI,CAAE,CAAW,EAC5D,CACF,EACA,GAAI,CAAC,OAAO,SAAS,CAAK,EACxB,MAAO,CACL,aAAc,EAAa,aAC3B,iBAAkB,aAAa,EAAQ,OAAO,MAAM,0BACtD,EAEF,IAAM,EACJ,EAAQ,kBAAoB,IAAA,KAC3B,EAAQ,OAAO,eAAiB,GAAS,EAAQ,gBAAkB,GAAS,EAAQ,iBACvF,MAAO,CACL,aAAc,EAAa,EAAa,SAAW,EAAa,aAChE,QACA,WAAY,EAAQ,OAAO,MAC3B,iBAAkB,EAAmB,EAAO,EAAa,OAAQ,EAAY,CAAO,CACtF,CACF,CAEA,SAAS,EACP,EACA,EACA,EACA,EACQ,CACR,GAAM,CAAE,QAAO,kBAAmB,EAAQ,OACpC,EAAY,EAA4B,EAAO,EAAQ,eAAe,EACtE,EACJ,EAAQ,kBAAoB,IAAA,GACxB,GACA,WAAW,EAAM,GAAG,EAAiB,IAAM,IAAI,GAAG,EAAY,EAAQ,gBAAiB,CAAS,EAAE,GAAG,EAAa,KAAO,MAAM,GACrI,MAAO;;IAEL,EAAM,KAAK,EAAY,EAAO,CAAS,EAAE;;QAErC,EAAS,eAAe,OAAO,EAAE,GAAG,EAAe;CAE3D,CAMA,SAAS,EAA4B,EAAe,EAAuC,CAEzF,GAAI,IAAc,IAAA,IAAa,IAAU,EAAW,MAAO,GAC3D,IAAK,IAAI,EAAY,EAAkB,EAAY,GAAI,IACrD,GAAI,EAAY,EAAO,CAAS,IAAM,EAAY,EAAW,CAAS,EAAG,OAAO,EAElF,MAAO,GACT,CAEA,SAAS,EAAY,EAAe,EAA2B,CAC7D,OAAO,OAAO,EAAM,YAAY,CAAS,CAAC,CAAC,CAAC,SAAS,CACvD,CAEA,SAAS,EACP,EACA,EACA,EACA,EAC8B,CAC9B,IAAI,EACJ,GAAI,CACF,EAAU,EAAgB,CAAc,CAC1C,OAAS,EAAO,CAGd,MAAO,KAAK,EAAmB,sBAD7B,aAAiB,EAAgB,GAAG,EAAM,WAAW,QAAQ,EAAyB,EAAM,UAAY,IAE5G,CACA,IAAM,EAAU,EAAY,EAAQ,EAAE,EAAE,OAAS,CAAC,EAAG,CAAO,EAC5D,GAAI,CAAC,EACH,MAAO,KAAK,EAAmB,YAAY,EAAQ,SAAS,QAAQ,EAAQ,aAAa,gBAG3F,IAAM,EAAc,IAAI,IAClB,EAAqB,CAAC,EACxB,EAAe,EACb,EAAiB,GAA0B,CAC/C,IACI,EAAS,OAAS,IAAgB,EAAS,KAAK,CAAO,CAC7D,EACA,IAAK,GAAM,CAAE,MAAO,EAAK,gBAAgB,EAAQ,MAAM,CAAC,EAAG,CACzD,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAa,EAAI,EAAQ,cAAgB,GAC/C,GAAI,CAAC,EAAO,IAAI,CAAE,EAChB,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,iBAAiB,OACjF,GAAI,EAAY,IAAI,CAAE,EAC3B,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,aAAa,MAC7E,CACL,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAU,EAClE,GACF,EAAc,GAAG,EAAW,QAAQ,EAAQ,aAAa,QAAQ,EAAK,CAAU,EAAE,KAAK,GAAiB,EAE1G,EAAY,IAAI,EAAI,CAAU,CAChC,CACF,CACA,IAAM,EAAa,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,OAAQ,GAAO,CAAC,EAAY,IAAI,CAAE,CAAC,EAczE,OAbI,EAAW,OAAS,IAEtB,IACA,EAAS,QACP,KAAK,EAAQ,SAAS,KAAK,EAAQ,CAAU,EAAE,cAAc,EAAW,OAAO,eAAe,OAAO,EAAE,GACzG,GAEE,EAAe,EACV,KAAK,EAAmB;;EAEjC,EAAS,IAAK,GAAY,KAAK,GAAS,CAAC,CAAC,KAAK;CAAI,IAAI,EAAe,EAAS,OAAS,UAAU,EAAe,EAAS,OAAA,CAAQ,eAAe,OAAO,EAAE,GAAK,GAAG;EAG3J,CACT,CAEA,eAAe,EAAW,EAAwB,EAAqE,CACrH,IAAM,EAAO,EAAS,MAAM,EAAG,SAAS,EAAgB,MAAM,CAAC,EACzD,EAAU,EAAY,EAAK,IAAM,CAAC,EAAG,CAAO,EAClD,GAAI,CAAC,EACH,MAAU,MAAM,yBAAyB,EAAQ,SAAS,OAAO,EAAQ,aAAa,YAAY,GAAgB,EAEpH,IAAM,EAAS,IAAI,IACnB,IAAK,IAAM,KAAO,EAAK,MAAM,CAAC,EAAG,CAC/B,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAQ,EAAI,EAAQ,cAAgB,GAC1C,GAAI,CAAC,EAAI,MAAU,MAAM,SAAS,EAAQ,SAAS,mBAAmB,GAAgB,EACtF,GAAI,EAAO,IAAI,CAAE,EAAG,MAAU,MAAM,aAAa,EAAQ,SAAS,mBAAmB,GAAI,EACzF,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAK,EACjE,GAAI,EACF,MAAU,MAAM,WAAW,EAAQ,aAAa,sBAAsB,EAAG,IAAI,GAAiB,EAChG,EAAO,IAAI,EAAI,CAAK,CACtB,CACA,GAAI,EAAO,OAAS,EAAG,MAAU,MAAM,4BAA4B,GAAgB,EACnF,OAAO,CACT,CAEA,SAAS,EACP,EACA,EACsD,CACtD,IAAM,EAAQ,EAAO,IAAK,GAAS,EAAK,KAAK,CAAC,EACxC,EAAU,EAAM,QAAQ,EAAQ,QAAQ,EACxC,EAAc,EAAM,QAAQ,EAAQ,YAAY,EACtD,OAAO,IAAY,IAAM,IAAgB,GAAK,IAAA,GAAY,CAAE,UAAS,aAAY,CACnF,CAEA,SAAS,EAAW,EAAiC,CACnD,OAAO,EAAI,MAAO,GAAS,EAAK,KAAK,IAAM,EAAE,CAC/C,CAEA,SAAS,EAAQ,EAAgC,CAC/C,IAAM,EAAS,EAAI,MAAM,EAAG,EAAc,CAAC,CAAC,IAAK,GAAO,KAAK,EAAK,CAAE,EAAE,GAAG,EACzE,OAAO,EAAI,OAAS,GAAiB,GAAG,EAAO,KAAK,IAAI,EAAE,OAAS,EAAO,KAAK,IAAI,CACrF,CAEA,SAAS,EAAK,EAAuB,CAEnC,IAAM,EAAY,EAAM,WAAW,aAAc,GAAG,CAAC,CAAC,WAAW,IAAK,GAAG,EAEzE,OADI,EAAU,KAAK,IAAM,GAAW,MAC7B,EAAU,OAAS,GAAyB,GAAG,EAAU,MAAM,EAAG,EAAsB,EAAE,GAAK,CACxG,CAEA,SAAS,EAAsB,EAA0B,EAAuD,CAC9G,OAAQ,EAAO,eAAiB,EAAa,WAAe,EAAY,iBAAmB,WAC7F,CAEA,SAAS,EAAqB,EAAwC,CACpE,OAAO,OAAO,SAAS,OAAO,CAAU,CAAC,GAAK,EAAW,KAAK,IAAM,GAAK,IAAA,GAAY,WACvF,CAEA,SAAS,EAAK,EAAmC,CAC/C,OAAO,EAAO,QAAQ,EAAK,IAAU,EAAM,EAAO,CAAC,EAAI,EAAO,MAChE"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@exercode/problem-utils",
-  "version": "2.3.0",
+  "version": "2.3.1",
   "description": ":100: A set of utilities for judging programs on Exercode (https://exercode.willbooster.com/).",
   "keywords": [
     "exercode",