evals api mvp

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

@@ -35,8 +35,7 @@ class MetaReferenceEvalsImpl(Evals):
         task: str,
     ) -> EvaluateResponse:
         cprint(f"model={model}, dataset={dataset}, task={task}", "red")
-        dataset = get_dataset("mmlu-simple-eval-en")
-
+        dataset = get_dataset(dataset)
         task_impl = get_task(task, dataset)
         x1 = task_impl.preprocess()
 
@@ -52,11 +51,6 @@ class MetaReferenceEvalsImpl(Evals):
                 generation_outputs.append(x.completion_message.content)
 
         x2 = task_impl.postprocess(generation_outputs)
-        scores = task_impl.score(x2)
-        print(scores)
-
-        return EvaluateResponse(
-            metrics={
-                "accuracy": 0.5,
-            }
-        )
+        eval_results = task_impl.score(x2)
+        eval_response = task_impl.aggregate_results(eval_results)
+        return eval_response

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

@@ -6,6 +6,7 @@
 import re
 
 from .task import BaseTask
+from llama_stack.apis.evals import *  # noqa: F403
 
 QUERY_TEMPLATE_MULTICHOICE = """
 Answer the following multiple choice question and make the answer very simple. The last line of your response should be of the following format: 'Answer: $LETTER' (without quotes) where LETTER is one of ABCD.
@@ -119,9 +120,6 @@ class MMLUTask(BaseTask):
         super().__init__(dataset, *args, **kwargs)
 
     def preprocess_sample(self, sample):
-        """
-        F1: preprocess sample
-        """
         content = QUERY_TEMPLATE_MULTICHOICE.format(**sample)
         return {
             "role": "user",
@@ -129,16 +127,10 @@ class MMLUTask(BaseTask):
         }
 
     def postprocess_sample(self, sample):
-        """
-        F2: postprocess sample
-        """
         normalized = normalize_response(sample)
         return normalized
 
     def score_sample(self, sample, expected):
-        """
-        F3: score sample
-        """
         extracted_answer = None
         for answer_regex in MULTILINGUAL_ANSWER_REGEXES:
             regex = MULTILINGUAL_ANSWER_PATTERN_TEMPLATE.format(answer_regex)
@@ -149,4 +141,10 @@ class MMLUTask(BaseTask):
         score = (
             1.0 if extracted_answer and extracted_answer == expected["Answer"] else 0.0
         )
+        # TODO: generalize this into SingleEvalResult
         return score
+
+    def aggregate_results(self, eval_results):
+        return EvaluateResponse(
+            metrics={"score": sum(eval_results) / len(eval_results)}
+        )

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

@@ -21,23 +21,18 @@ class BaseTask(ABC):
 
     @abstractmethod
     def preprocess_sample(self, sample):
-        """
-        F1: preprocess sample
-        """
         raise NotImplementedError()
 
     @abstractmethod
     def postprocess_sample(self, sample):
-        """
-        F2: postprocess sample
-        """
         raise NotImplementedError()
 
     @abstractmethod
     def score_sample(self, sample, ground_truth):
-        """
-        F3: score sample
-        """
+        raise NotImplementedError()
+
+    @abstractmethod
+    def aggregate_results(self, eval_results):
         raise NotImplementedError()
 
     def preprocess(self):