@m4trix/evals 0.20.0 → 0.21.0

package/dist/index.d.ts

@@ -437,6 +437,15 @@ type RunnerEvent = {
     type: 'RunStarted';
     runId: string;
     startedAt: number;
+} | {
+    type: 'TestCaseStarted';
+    runId: string;
+    testCaseId: string;
+    testCaseName: string;
+    startedTestCases: number;
+    totalTestCases: number;
+    rerunIndex: number;
+    rerunTotal: number;
 } | {
     type: 'TestCaseProgress';
     runId: string;

package/dist/index.js

@@ -1106,13 +1106,27 @@ function createArtifactPath(artifactDirectory, datasetId, runId) {
     `${datasetId}_${runId}_${nowIsoForFile()}.jsonl`
   );
 }
-function processOneTestCase(task, testCaseItem, totalEvaluations, publishEvent, persistenceQueue, updateSnapshot, completedRef, passedRef, failedRef) {
+function processOneTestCase(task, testCaseItem, totalEvaluations, publishEvent, persistenceQueue, updateSnapshot, startedRef, completedRef, passedRef, failedRef) {
   return Effect.gen(function* () {
     const reruns = typeof testCaseItem.testCase.getReruns === "function" ? testCaseItem.testCase.getReruns() : 1;
     const rerunPassed = [];
     for (let r = 0; r < reruns; r++) {
       const evaluatorRunId = `run-${randomUUID()}`;
       const started = Date.now();
+      const startedEvaluations = yield* Ref.modify(startedRef, (n) => [
+        n + 1,
+        n + 1
+      ]);
+      yield* publishEvent({
+        type: "TestCaseStarted",
+        runId: task.runId,
+        testCaseId: testCaseItem.id,
+        testCaseName: testCaseItem.testCase.getName(),
+        startedTestCases: startedEvaluations,
+        totalTestCases: totalEvaluations,
+        rerunIndex: r + 1,
+        rerunTotal: reruns
+      });
       const evaluatorScores = [];
       let testCaseError;
       const output = readOutput(testCaseItem.testCase);
@@ -1258,6 +1272,7 @@ var executeRunTask = (task, publishEvent, persistenceQueue, updateSnapshot) => E
   );
   const maxConcurrency = Math.max(1, task.maxConcurrency ?? 1);
   const completedRef = yield* Ref.make(0);
+  const startedRef = yield* Ref.make(0);
   const passedRef = yield* Ref.make(0);
   const failedRef = yield* Ref.make(0);
   const processTestCase = (testCaseItem) => processOneTestCase(
@@ -1267,6 +1282,7 @@ var executeRunTask = (task, publishEvent, persistenceQueue, updateSnapshot) => E
     publishEvent,
     persistenceQueue,
     updateSnapshot,
+    startedRef,
     completedRef,
     passedRef,
     failedRef