@m4trix/evals 0.18.0 → 0.20.0

package/README.md

@@ -77,8 +77,15 @@ export const myEvaluator = Evaluator.define({
   inputSchema,
   outputSchema: S.Unknown,
   scoreSchema: S.Struct({ scores: S.Array(S.Unknown) }),
-}).evaluate(async ({ input, ctx: _ctx, output }) => {
+}).evaluate(async ({ input, ctx: _ctx, output, createError }) => {
   const start = Date.now();
+  const value = 85;
+  if (value < 50) {
+    return createError(
+      { reason: 'score below minimum', value, prompt: input.prompt, output },
+      { label: 'quality-check' },
+    );
+  }
   const latencyMs = Date.now() - start;
   const minScore =
     typeof output === 'object' &&
@@ -90,7 +97,7 @@ export const myEvaluator = Evaluator.define({
   return {
     scores: [
       percentScore.make(
-        { value: 85 },
+        { value },
         { definePassed: (d) => d.value >= (minScore ?? 50) },
       ),
     ],

package/dist/cli-simple.cjs

@@ -294,6 +294,8 @@ function createDiffString(expected, actual, diffOptions) {
 function formatLogMessage(msg) {
   if (typeof msg === "string")
     return msg;
+  if (msg instanceof Error)
+    return msg.stack ?? msg.message;
   try {
     if (msg !== null && typeof msg === "object") {
       return JSON.stringify(msg, null, 2);
@@ -347,7 +349,11 @@ var Metric = {
       name: config.name,
       aggregate: config.aggregate,
       format: config.format,
-      make: (data) => ({ id: config.id, data })
+      make: (data, options) => ({
+        id: config.id,
+        data,
+        ...options?.name !== void 0 && { name: options.name }
+      })
     };
     registry.set(config.id, def);
     return def;
@@ -369,25 +375,61 @@ var ScoreAggregate = {
       const count = values.length || 1;
       const result = {};
       for (const field of fields) {
-        result[field] = values.reduce((s, v) => s + (v[field] ?? 0), 0) / count;
+        result[field] = values.reduce(
+          (s, v) => s + (v[field] ?? 0),
+          0
+        ) / count;
       }
       return result;
     };
   },
-  /** Average `value` with sample std dev. Use for percent-style scores. */
-  averageWithVariance(values) {
-    if (values.length === 0) {
-      return { value: 0, stdDev: void 0, count: 0 };
-    }
-    const sum = values.reduce((s, v) => s + v.value, 0);
-    const sumSq = values.reduce((s, v) => s + v.value * v.value, 0);
-    const mean = sum / values.length;
-    let stdDev;
-    if (values.length >= 2) {
-      const variance = (sumSq - values.length * mean * mean) / (values.length - 1);
-      stdDev = variance > 0 ? Math.sqrt(variance) : 0;
-    }
-    return { ...values[0], value: mean, stdDev, count: values.length };
+  /** Average selected numeric fields, with sample std dev tracked for `value`. */
+  averageWithVariance(fields) {
+    return (values) => {
+      const count = values.length;
+      const result = {};
+      for (const field of fields) {
+        result[field] = count === 0 ? 0 : values.reduce(
+          (sum, item) => sum + (item[field] ?? 0),
+          0
+        ) / count;
+      }
+      const valueField = "value";
+      const hasValueField = fields.includes(valueField);
+      if (count === 0) {
+        if (hasValueField) {
+          result[valueField] = 0;
+        }
+        return {
+          ...result,
+          stdDev: void 0,
+          count: 0
+        };
+      }
+      let stdDev;
+      if (hasValueField && count >= 2) {
+        const sum = values.reduce(
+          (s, v) => s + (v[valueField] ?? 0),
+          0
+        );
+        const sumSq = values.reduce(
+          (s, v) => {
+            const value = v[valueField] ?? 0;
+            return s + value * value;
+          },
+          0
+        );
+        const mean = sum / count;
+        const variance = (sumSq - count * mean * mean) / (count - 1);
+        stdDev = variance > 0 ? Math.sqrt(variance) : 0;
+      }
+      return {
+        ...values[0],
+        ...result,
+        stdDev,
+        count
+      };
+    };
   },
   /** All runs must pass. Use for binary scores. */
   all(values) {
@@ -421,6 +463,7 @@ var Score = {
           id: config.id,
           data,
           ...passed !== void 0 && { passed },
+          ...options?.name !== void 0 && { name: options.name },
           def
           // Attach def so rendering/aggregation works without registry lookup
         };
@@ -489,7 +532,7 @@ Score.of({
   displayStrategy: "bar",
   formatValue: (data) => data.value.toFixed(2),
   formatAggregate: (data) => data.stdDev != null ? `Avg: ${data.value.toFixed(2)} \xB1 ${data.stdDev.toFixed(2)}` : `Avg: ${data.value.toFixed(2)}`,
-  aggregateValues: Score.aggregate.averageWithVariance
+  aggregateValues: Score.aggregate.averageWithVariance(["value"])
 });
 Score.of({
   id: "delta",
@@ -518,6 +561,14 @@ Score.of({
 function getScoreDef(item) {
   return item.def ?? getScoreById(item.id);
 }
+function lastNonEmptyName(items) {
+  for (let i = items.length - 1; i >= 0; i--) {
+    const n = items[i].name;
+    if (n != null && n.trim().length > 0)
+      return n;
+  }
+  return void 0;
+}
 function aggregateScoreItems(items) {
   if (items.length === 0)
     return void 0;
@@ -525,7 +576,13 @@ function aggregateScoreItems(items) {
   if (!def?.aggregateValues)
     return items[items.length - 1];
   const aggregated = def.aggregateValues(items.map((i) => i.data));
-  return { ...items[0], data: aggregated, def };
+  const nameOverride = lastNonEmptyName(items);
+  return {
+    ...items[0],
+    data: aggregated,
+    def,
+    ...nameOverride !== void 0 && { name: nameOverride }
+  };
 }
 function aggregateMetricItems(items) {
   if (items.length === 0)
@@ -534,7 +591,12 @@ function aggregateMetricItems(items) {
   if (!def?.aggregate)
     return items[items.length - 1];
   const aggregated = def.aggregate(items.map((i) => i.data));
-  return { ...items[0], data: aggregated };
+  const nameOverride = lastNonEmptyName(items);
+  return {
+    ...items[0],
+    data: aggregated,
+    ...nameOverride !== void 0 && { name: nameOverride }
+  };
 }
 function toNumericScoreFromScores(scores) {
   for (const item of scores) {
@@ -573,6 +635,7 @@ function toNumericScore(value) {
 }
 
 // src/runner/execution.ts
+var evaluatorErrorLogEntryKey = "__m4trixEvaluatorLogEntry";
 function computeEvaluatorPassed(evaluator, result, scores) {
   const scoresWithPassed = scores.filter((s) => "passed" in s && s.passed !== void 0);
   if (scoresWithPassed.length > 0) {
@@ -629,20 +692,26 @@ function processOneTestCase(task, testCaseItem, totalEvaluations, publishEvent,
         if (!evaluateFn) {
           continue;
         }
+        const logs = [];
+        const logDiff = (expected, actual, options) => {
+          logs.push(createDiffLogEntry(expected, actual, options));
+        };
+        const log = (message, options) => {
+          logs.push(createLogEntry(message, options));
+        };
+        const createError = (message, options) => {
+          const entry = createLogEntry(message, options);
+          const error = message instanceof Error ? message : new Error(entry.message);
+          error[evaluatorErrorLogEntryKey] = entry;
+          return error;
+        };
         try {
-          const logs = [];
-          const logDiff = (expected, actual, options) => {
-            logs.push(createDiffLogEntry(expected, actual, options));
-          };
-          const log = (message, options) => {
-            logs.push(createLogEntry(message, options));
-          };
           const ctx = yield* effect.Effect.promise(
             () => Promise.resolve(evaluator.resolveContext())
           );
           const result = yield* effect.Effect.promise(
-            () => Promise.resolve(
-              evaluateFn({
+            () => Promise.resolve().then(
+              () => evaluateFn({
                 input: testCaseItem.testCase.getInput(),
                 ctx,
                 output,
@@ -652,10 +721,24 @@ function processOneTestCase(task, testCaseItem, totalEvaluations, publishEvent,
                   datasetId: task.datasetId
                 },
                 logDiff,
-                log
+                log,
+                createError
               })
             )
           );
+          if (result instanceof Error) {
+            const evaluatorError = result;
+            const taggedEntry = evaluatorError[evaluatorErrorLogEntryKey];
+            logs.push(taggedEntry ?? createLogEntry(result));
+            testCaseError = result.message;
+            evaluatorScores.push({
+              evaluatorId,
+              scores: [],
+              passed: false,
+              logs: logs.length > 0 ? logs : void 0
+            });
+            continue;
+          }
           const { scores, metrics } = normalizeResult(result);
           const passed2 = computeEvaluatorPassed(evaluator, result, scores);
           evaluatorScores.push({
@@ -666,11 +749,16 @@ function processOneTestCase(task, testCaseItem, totalEvaluations, publishEvent,
             logs: logs.length > 0 ? logs : void 0
           });
         } catch (error) {
+          if (error instanceof Error) {
+            const taggedEntry = error[evaluatorErrorLogEntryKey];
+            logs.push(taggedEntry ?? createLogEntry(error));
+          }
           testCaseError = error instanceof Error ? error.message : "Evaluator execution failed";
           evaluatorScores.push({
             evaluatorId,
             scores: [],
-            passed: false
+            passed: false,
+            logs: logs.length > 0 ? logs : void 0
           });
         }
       }
@@ -1654,6 +1742,7 @@ function RunView({
               rerunTotal: event.rerunTotal,
               durationMs: events.reduce((s, e) => s + e.durationMs, 0),
               passed: events.every((e) => e.passed),
+              errorMessage: event.errorMessage,
               events,
               aggregatedEvaluatorScores,
               isAggregated
@@ -1764,8 +1853,13 @@ function RunView({
           " (",
           tc.durationMs,
           "ms)"
-        ] })
+        ] }),
+        tc.errorMessage ? /* @__PURE__ */ jsxRuntime.jsxs(ink.Text, { color: "red", bold: true, children: [
+          " ",
+          "ERROR"
+        ] }) : null
       ] }),
+      tc.errorMessage ? /* @__PURE__ */ jsxRuntime.jsx(ink.Text, { color: "red", children: tc.errorMessage }) : null,
       tc.aggregatedEvaluatorScores.map((item) => /* @__PURE__ */ jsxRuntime.jsxs(
         ink.Box,
         {
@@ -1786,9 +1880,10 @@ function RunView({
                   const formatted = def.format(m.data, {
                     isAggregated: tc.isAggregated
                   });
+                  const label = m.name ?? def.name;
                   return /* @__PURE__ */ jsxRuntime.jsxs(ink.Text, { color: "gray", children: [
                     "[",
-                    def.name ? `${def.name}: ` : "",
+                    label ? `${label}: ` : "",
                     formatted,
                     "]",
                     " "
@@ -1797,8 +1892,8 @@ function RunView({
               ] }) : null
             ] }),
             item.scores.length > 0 ? item.scores.map((s, idx) => {
-              const def = getScoreById(s.id);
-              const scoreLabel = def ? def.name ?? def.id : s.id;
+              const def = s.def ?? getScoreById(s.id);
+              const scoreLabel = s.name ?? def?.name ?? def?.id ?? s.id;
               return /* @__PURE__ */ jsxRuntime.jsxs(
                 ink.Text,
                 {
@@ -1902,7 +1997,7 @@ function RunView({
               if (!aggregated)
                 return null;
               const def = aggregated.def ?? getScoreById(aggregated.id);
-              const label = def ? def.name ?? def.id : aggregated.id;
+              const label = aggregated.name ?? def?.name ?? def?.id ?? aggregated.id;
               const formatted = def ? def.formatAggregate(aggregated.data) : "n/a";
               const numeric = toNumericScore(aggregated.data);
               return /* @__PURE__ */ jsxRuntime.jsxs(
@@ -2063,7 +2158,7 @@ function getEvaluatorSummaryLines(evaluatorId, evaluatorName, aggregate, scoreIt
     if (!agg)
       continue;
     const def = agg.def ?? getScoreById(agg.id);
-    const label = def ? def.name ?? def.id : agg.id;
+    const label = agg.name ?? def?.name ?? def?.id ?? agg.id;
     const formatted = def ? def.formatAggregate(agg.data) : "n/a";
     const numeric = toNumericScore(agg.data);
     const colored = numeric !== void 0 ? colorize(formatted, scoreToColor(numeric)) : formatted;
@@ -2129,12 +2224,13 @@ function formatEvaluatorScoreLine(name, scores, passed, metrics, options) {
   const passLabel = passed ? colorize("PASS", `${ansi2.bold}${ansi2.green}`) : colorize("FAIL", `${ansi2.bold}${ansi2.red}`);
   const metricParts = [];
   if (metrics && metrics.length > 0) {
-    for (const { id, data } of metrics) {
-      const def = getMetricById(id);
+    for (const m of metrics) {
+      const def = getMetricById(m.id);
       if (def) {
-        const formatted = def.format(data, options);
+        const formatted = def.format(m.data, options);
+        const label = m.name ?? def.name;
         metricParts.push(
-          def.name ? `[${def.name}: ${formatted}]` : `[${formatted}]`
+          label ? `[${label}: ${formatted}]` : `[${formatted}]`
         );
       }
     }
@@ -2142,7 +2238,7 @@ function formatEvaluatorScoreLine(name, scores, passed, metrics, options) {
   const scoreLines = [];
   for (const item of scores) {
     const def = item.def ?? getScoreById(item.id);
-    const scoreLabel = def ? def.name ?? def.id : item.id;
+    const scoreLabel = item.name ?? def?.name ?? def?.id ?? item.id;
     let formatted;
     if (!def) {
       const numeric = toNumericScore(item.data);
@@ -2301,9 +2397,13 @@ async function runSimpleEvalCommandPlain(runner, datasetName, evaluatorPattern)
           0
         );
         const lines = [];
+        const statusSuffix = event.errorMessage ? ` ${colorize("ERROR", `${ansi2.bold}${ansi2.red}`)}` : "";
         lines.push(
-          `${colorize(`[${event.completedTestCases}/${event.totalTestCases}]`, ansi2.cyan)} ${event.testCaseName} ${colorize(`(${event.rerunIndex}/${event.rerunTotal})`, ansi2.cyan)} ${colorize(`(${durationMs}ms)`, ansi2.dim)}`
+          `${colorize(`[${event.completedTestCases}/${event.totalTestCases}]`, ansi2.cyan)} ${event.testCaseName} ${colorize(`(${event.rerunIndex}/${event.rerunTotal})`, ansi2.cyan)} ${colorize(`(${durationMs}ms)`, ansi2.dim)}${statusSuffix}`
         );
+        if (event.errorMessage) {
+          lines.push(colorize(event.errorMessage, ansi2.red));
+        }
         for (const item of aggregatedScores) {
           const name = evaluatorNameById.get(item.evaluatorId) ?? item.evaluatorId;
           lines.push(