( judge: Judge, systemPrompt: string, diffPair: DiffPair, scoreType: "logic-equivalence", runs: number = 3, )
| 55 | } |
| 56 | |
| 57 | async function testConsistency( |
| 58 | judge: Judge, |
| 59 | systemPrompt: string, |
| 60 | diffPair: DiffPair, |
| 61 | scoreType: "logic-equivalence", |
| 62 | runs: number = 3, |
| 63 | ): Promise<ScoreResult[]> { |
| 64 | const userPrompt = createDiffComparisonPrompt(diffPair, scoreType); |
| 65 | |
| 66 | const results = await Promise.all( |
| 67 | Array.from({ length: runs }, () => |
| 68 | evaluateWithJudge(judge, systemPrompt, userPrompt), |
| 69 | ), |
| 70 | ); |
| 71 | |
| 72 | return results; |
| 73 | } |
| 74 | |
| 75 | const testJudge = judges[0]; |
| 76 |
no test coverage detected