scorer registry

llama_stack/distribution/registry/scorers/__init__.py

@@ -5,9 +5,19 @@
 # the root directory of this source tree.
 # TODO: make these import config based
 from llama_stack.apis.evals import *  # noqa: F403
+from llama_stack.providers.impls.meta_reference.evals.scorer.basic_scorers import *  # noqa: F403
 
 from ..registry import Registry
 
 
 class ScorerRegistry(Registry[BaseScorer]):
     _REGISTRY: Dict[str, BaseScorer] = {}
+
+
+SCORER_REGISTRY = {
+    "accuracy": AccuracyScorer,
+    "random": RandomScorer,
+}
+
+for k, v in SCORER_REGISTRY.items():
+    ScorerRegistry.register(k, v)

llama_stack/providers/impls/meta_reference/evals/evals.py

@@ -53,6 +53,7 @@ class MetaReferenceEvalsImpl(Evals):
                 scoring_config=EvaluateScoringConfig(
                     scorer_config_list=[
                         EvaluateSingleScorerConfig(scorer_name="accuracy"),
+                        EvaluateSingleScorerConfig(scorer_name="random"),
                     ]
                 ),
             )

llama_stack/providers/impls/meta_reference/evals/scorer/aggregate_scorer.py

@@ -0,0 +1,35 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+from llama_stack.apis.evals.evals import BaseScorer, EvalResult, SingleEvalResult
+from llama_stack.apis.datasets.datasets import *  # noqa: F401 F403
+
+
+class AggregateScorer(BaseScorer[ScorerInputSample]):
+    def __init__(self, scorers: List[BaseScorer[ScorerInputSample]]):
+        self.scorers = scorers
+
+    def score_sample(self, scorer_input_sample: ScorerInputSample) -> SingleEvalResult:
+        all_score_data = {}
+        for scorer in self.scorers:
+            score_data = scorer.score_sample(scorer_input_sample).score_data
+            for k, v in score_data.items():
+                all_score_data[k] = v
+
+        return SingleEvalResult(
+            score_data=all_score_data,
+        )
+
+    def aggregate_results(self, eval_results: List[SingleEvalResult]) -> EvalResult:
+        all_metrics = {}
+
+        for scorer in self.scorers:
+            metrics = scorer.aggregate_results(eval_results).metrics
+            for k, v in metrics.items():
+                all_metrics[f"{scorer.__class__.__name__}:{k}"] = v
+
+        return EvalResult(
+            metrics=all_metrics,
+        )

llama_stack/providers/impls/meta_reference/evals/scorer/basic_scorers.py

@@ -9,34 +9,6 @@ from llama_stack.apis.evals.evals import BaseScorer, EvalResult, SingleEvalResul
 from llama_stack.apis.datasets.datasets import *  # noqa: F401 F403
 
 
-class AggregateScorer(BaseScorer[ScorerInputSample]):
-    def __init__(self, scorers: List[BaseScorer[ScorerInputSample]]):
-        self.scorers = scorers
-
-    def score_sample(self, scorer_input_sample: ScorerInputSample) -> SingleEvalResult:
-        all_score_data = {}
-        for scorer in self.scorers:
-            score_data = scorer.score_sample(scorer_input_sample).score_data
-            for k, v in score_data.items():
-                all_score_data[k] = v
-
-        return SingleEvalResult(
-            score_data=all_score_data,
-        )
-
-    def aggregate_results(self, eval_results: List[SingleEvalResult]) -> EvalResult:
-        all_metrics = {}
-
-        for scorer in self.scorers:
-            metrics = scorer.aggregate_results(eval_results).metrics
-            for k, v in metrics.items():
-                all_metrics[f"{scorer.__class__.__name__}:{k}"] = v
-
-        return EvalResult(
-            metrics=all_metrics,
-        )
-
-
 class RandomScorer(BaseScorer[ScorerInputSample]):
     def score_sample(self, scorer_input_sample: ScorerInputSample) -> SingleEvalResult:
         return SingleEvalResult(score_data={"random": random.random()})

llama_stack/providers/impls/meta_reference/evals/tasks/run_eval_task.py

@@ -4,6 +4,8 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 from llama_stack.distribution.registry.datasets import DatasetRegistry
+from llama_stack.distribution.registry.scorers import ScorerRegistry
+from llama_stack.providers.impls.meta_reference.evals.scorer.aggregate_scorer import *  # noqa: F403
 from llama_stack.providers.impls.meta_reference.evals.scorer.basic_scorers import *  # noqa: F403
 from llama_stack.providers.impls.meta_reference.evals.generator.inference_generator import (
     InferenceGenerator,
@@ -59,11 +61,14 @@ class RunEvalTask(BaseTask):
         cprint(postprocessed, "blue")
 
         # F3 - scorer
+        scorer_config_list = eval_task_config.scoring_config.scorer_config_list
+        scorer_list = []
+        for s_conf in scorer_config_list:
+            scorer = ScorerRegistry.get(s_conf.scorer_name)
+            scorer_list.append(scorer())
+
         scorer = AggregateScorer(
-            scorers=[
+            scorers=scorer_list,
-                AccuracyScorer(),
-                RandomScorer(),
-            ]
         )
 
         scorer_results = scorer.score(postprocessed)