| 1 | export interface CredibilityIssue { |
| 2 | line: number |
| 3 | text: string |
| 4 | reason: string |
| 5 | } |
| 6 | |
| 7 | const BENCHMARK_PATTERN = /\b(MMLU|HumanEval|MT-Bench|SWE-bench|GSM8K|GPQA|BIG-bench)\b/i |
| 8 | const METRIC_WORD_PATTERN = |
| 9 | /成本|价格|准确率|召回率|错误率|幻觉|提升|下降|降低|缩短|节省|突破|达到|保持|超过|优化|损失|参数|token|fps|benchmark|score|accuracy|cost|latency|throughput/i |
| 10 | const EXACT_VALUE_PATTERN = |
| 11 | /(?:[$￥]\s*\d|\d+(?:\.\d+)?\s*(?:%|美元|元|倍|fps|K|M|B|T|万|亿|千亿|万亿|token|tokens?)|\b\d+(?:\.\d+)?\s*(?:->|→|至|-)\s*\d+(?:\.\d+)?)/i |
| 12 | |
| 13 | export function findUnsupportedPrecisionClaims(args: { |
| 14 | markdown: string |
| 15 | hasSources: boolean |
| 16 | }): CredibilityIssue[] { |
| 17 | if (args.hasSources) return [] |
| 18 | |
| 19 | const issues: CredibilityIssue[] = [] |
| 20 | const lines = args.markdown.split('\n') |
| 21 | |
| 22 | lines.forEach((line, index) => { |
| 23 | const text = line.trim() |
| 24 | if (!text || isAllowedStructuralNumber(text)) return |
| 25 | |
| 26 | const hasExactValue = EXACT_VALUE_PATTERN.test(text) |
| 27 | const hasBenchmark = BENCHMARK_PATTERN.test(text) && /\d/.test(text) |
| 28 | const hasMetricWord = METRIC_WORD_PATTERN.test(text) |
| 29 | |
| 30 | if ((hasExactValue && hasMetricWord) || hasBenchmark) { |
| 31 | issues.push({ |
| 32 | line: index + 1, |
| 33 | text, |
| 34 | reason: hasBenchmark |
| 35 | ? 'benchmark score without source support' |
| 36 | : 'exact metric without source support' |
| 37 | }) |
| 38 | } |
| 39 | }) |
| 40 | |
| 41 | return issues |
| 42 | } |
| 43 | |
| 44 | function isAllowedStructuralNumber(text: string): boolean { |
| 45 | return ( |
| 46 | /^##\s*Page\s+\d+\s*:/.test(text) || |
| 47 | /^##\s*Page Count$/i.test(text) || |
| 48 | /^\d+$/.test(text) || |
| 49 | /时长|分钟|页|页面数|Page Count/i.test(text) |
| 50 | ) |
| 51 | } |
| 52 |