generator + scorer Api for MMLU

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

@@ -3,16 +3,27 @@
 #
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
+import json
+
+from termcolor import cprint
+
+from llama_stack.providers.impls.meta_reference.evals.scorer.basic_scorers import (
+    AggregateScorer,
+)
 
 from llama_stack.apis.inference import *  # noqa: F403
 from llama_stack.apis.evals import *  # noqa: F403
 from llama_stack.apis.dataset import *  # noqa: F403
 
-from termcolor import cprint
-
 from llama_stack.distribution.registry.datasets.dataset_registry import DatasetRegistry
+from llama_stack.providers.impls.meta_reference.evals.processor.mmlu_processor import (
+    MMLUProcessor,
+)
+
+# from llama_stack.distribution.registry.tasks.task_registry import TaskRegistry
+# from .tasks.run_eval_task import RunEvalTask
+from .scorer.basic_scorers import *  # noqa: F403
 
-from llama_stack.distribution.registry.tasks.task_registry import TaskRegistry
 
 from .config import MetaReferenceEvalsImplConfig
 
@@ -27,7 +38,7 @@ class MetaReferenceEvalsImpl(Evals):
     async def shutdown(self) -> None:
         pass
 
-    async def run_evals(
+    async def run_eval_task(
         self,
         model: str,
         task: str,
@@ -38,43 +49,142 @@ class MetaReferenceEvalsImpl(Evals):
             f"model={model}, dataset={dataset}, task={task}, eval_task_config={eval_task_config}",
             "red",
         )
+
         if not dataset:
             raise ValueError("dataset must be specified for mete-reference evals")
 
-        dataset = DatasetRegistry.get_dataset(dataset)
-        dataset.load()
+        if not eval_task_config:
+            # construct eval task config from inputs
+            eval_task_config = EvaluateTaskConfig(
+                dataset_config=EvaluateDatasetConfig(
+                    dataset_name=dataset,
+                    row_limit=2,
+                ),
+                generation_config=EvaluateModelGenerationConfig(
+                    model=model,
+                ),
+                scoring_config=EvaluateScoringConfig(
+                    scorer_config_list=[
+                        EvaluateSingleScorerConfig(scorer_name="accuracy"),
+                    ]
+                ),
+            )
 
-        task_impl = TaskRegistry.get_task(task)()
-        preprocessed = task_impl.preprocess(dataset)
+        # TODO: wrap inside task
+        # run_task = RunEvalTask(
+        #     eval_task_config=eval_task_config,
+        # )
+        # eval_result = run_task.run()
 
-        # TODO: replace w/ batch inference & async return eval job
-        generation_outputs = []
-        if eval_task_config is None:
-            eval_task_config = EvaluateTaskConfig(n_samples=len(preprocessed))
-        if eval_task_config.n_samples is None or eval_task_config.n_samples > len(
-            preprocessed
-        ):
-            eval_task_config.n_samples = len(preprocessed)
-
-        print(
-            f"Eval generation start, generate on {eval_task_config.n_samples} samples"
+        dataset = DatasetRegistry.get_dataset(
+            eval_task_config.dataset_config.dataset_name
         )
+        dataset.load(n_samples=eval_task_config.dataset_config.row_limit)
+        print(f"Running on {len(dataset)} samples")
 
-        for sample in preprocessed[: eval_task_config.n_samples]:
+        # F1
+        processor = MMLUProcessor()
+        preprocessed = processor.preprocess(dataset)
+
+        # Generation
+        # TODO: wrap inside BaseGenerator
+        generation_outputs = []
+        for sample in preprocessed:
             print("generation: ", sample)
             response = await self.inference_api.chat_completion(
                 model=model,
-                messages=sample.preprocessed["messages"],
+                messages=sample.generation_input.messages,
                 stream=False,
             )
-            sample.prediction = PredictionSample(
-                completion_message=response.completion_message.content
-            )
-            generation_outputs.append(sample)
+            cprint(f"response: {response}", "cyan")
 
-        postprocessed = task_impl.postprocess(generation_outputs)
-        eval_results = task_impl.score(postprocessed)
-        aggr_result = task_impl.aggregate_results(eval_results)
-        return EvaluateResponse(
-            eval_result=aggr_result,
+            generation_outputs.append(
+                GenerationResponseSample(
+                    generation_output=GenerationOutput(
+                        completion_message=response.completion_message.content
+                    )
+                )
+            )
+        cprint(generation_outputs, "green")
+
+        # F2
+        postprocessed = processor.postprocess(generation_outputs, dataset)
+        cprint(postprocessed, "blue")
+
+        # F3 - scorer
+        scorer = AggregateScorer(
+            scorers=[
+                AccuracyScorer(),
+                RandomScorer(),
+            ]
         )
+
+        scorer_results = scorer.score(postprocessed)
+        cprint(scorer_results, "magenta")
+        eval_result = scorer.aggregate_results(scorer_results)
+
+        return EvaluateResponse(
+            eval_result=eval_result,
+            formatted_report=json.dumps(eval_result.json(), indent=4),
+        )
+
+    async def run_scorer(
+        self,
+        dataset_config: EvaluateDatasetConfig,
+        eval_scoring_config: EvaluateScoringConfig,
+    ) -> EvaluateResponse:
+        return EvaluateResponse(
+            eval_result={},
+        )
+
+    # async def run_evals(
+    #     self,
+    #     model: str,
+    #     task: str,
+    #     dataset: Optional[str] = None,
+    #     eval_task_config: Optional[EvaluateTaskConfig] = None,
+    # ) -> EvaluateResponse:
+    #     cprint(
+    #         f"model={model}, dataset={dataset}, task={task}, eval_task_config={eval_task_config}",
+    #         "red",
+    #     )
+    #     if not dataset:
+    #         raise ValueError("dataset must be specified for mete-reference evals")
+
+    #     dataset = DatasetRegistry.get_dataset(dataset)
+    #     dataset.load()
+
+    #     task_impl = TaskRegistry.get_task(task)()
+    #     preprocessed = task_impl.preprocess(dataset)
+
+    #     # TODO: replace w/ batch inference & async return eval job
+    #     generation_outputs = []
+    #     if eval_task_config is None:
+    #         eval_task_config = EvaluateTaskConfig(n_samples=len(preprocessed))
+    #     if eval_task_config.n_samples is None or eval_task_config.n_samples > len(
+    #         preprocessed
+    #     ):
+    #         eval_task_config.n_samples = len(preprocessed)
+
+    #     print(
+    #         f"Eval generation start, generate on {eval_task_config.n_samples} samples"
+    #     )
+
+    #     for sample in preprocessed[: eval_task_config.n_samples]:
+    #         print("generation: ", sample)
+    #         response = await self.inference_api.chat_completion(
+    #             model=model,
+    #             messages=sample.preprocessed["messages"],
+    #             stream=False,
+    #         )
+    #         sample.prediction = PredictionSample(
+    #             completion_message=response.completion_message.content
+    #         )
+    #         generation_outputs.append(sample)
+
+    #     postprocessed = task_impl.postprocess(generation_outputs)
+    #     eval_results = task_impl.score(postprocessed)
+    #     aggr_result = task_impl.aggregate_results(eval_results)
+    #     return EvaluateResponse(
+    #         eval_result=aggr_result,
+    #     )

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

@@ -0,0 +1,5 @@
+# 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.

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

@@ -7,8 +7,6 @@ import re
 
 from llama_stack.apis.evals import *  # noqa: F403
 
-# from llama_stack.distribution.registry.tasks.task import BaseTask
-
 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.
 
@@ -112,60 +110,78 @@ def normalize_extracted_answer(extracted_answer: str) -> str:
     )
 
 
-class MMLUTask(BaseTask[DictSample, ProcessedDictSample]):
+class MMLUProcessor(
+    BaseGeneratorProcessor[
+        DictSample, PreprocessedSample, GenerationResponseSample, ScorerInputSample
+    ]
+):
     """
-    MMLU Task.
+    Generator processor for MMLU
     """
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
 
-    def preprocess_sample(self, sample: ProcessedDictSample) -> ProcessedDictSample:
+    def preprocess_sample(self, sample: DictSample) -> PreprocessedSample:
         content = QUERY_TEMPLATE_MULTICHOICE.format(**sample.data)
-        preprocessed = {
-            "messages": [
-                {
-                    "role": "user",
-                    "content": content,
-                }
-            ],
-        }
-        processed_sample = ProcessedDictSample(
-            data=sample.data,
-            preprocessed=preprocessed,
+        preprocessed_msgs = [
+            {
+                "role": "user",
+                "content": content,
+            }
+        ]
+        processed_sample = PreprocessedSample(
+            generation_input=GenerationInput(
+                messages=preprocessed_msgs,
+            )
         )
         return processed_sample
 
-    def postprocess_sample(self, sample: ProcessedDictSample) -> ProcessedDictSample:
-        if not sample.postprocessed:
-            sample.postprocessed = {}
-        sample.postprocessed["postprocessed"] = normalize_response(
-            sample.prediction.completion_message
-        )
-        return sample
+    def postprocess_sample(
+        self, generation_sample: GenerationResponseSample, dataset_sample: DictSample
+    ) -> ScorerInputSample:
+        response_text = generation_sample.generation_output.completion_message
+        normalized_response = normalize_response(response_text)
 
-    def score_sample(self, sample: ProcessedDictSample) -> SingleEvalResult:
-        postprocessed_output = sample.postprocessed["postprocessed"]
-        expected_answer = sample.data["Answer"]
-
-        extracted_answer = None
+        # extract answer
+        extracted_answer = ""
         for answer_regex in MULTILINGUAL_ANSWER_REGEXES:
             regex = MULTILINGUAL_ANSWER_PATTERN_TEMPLATE.format(answer_regex)
-            match = re.search(regex, postprocessed_output)
+            match = re.search(regex, normalized_response)
             if match:
                 extracted_answer = normalize_extracted_answer(match.group(1))
                 break
 
-        score = 1.0 if extracted_answer and extracted_answer == expected_answer else 0.0
-
-        return SingleEvalResult(
-            score_data={
-                "score": score,
-            },
+        return ScorerInputSample(
+            generation_output=PostprocessedGeneration(
+                completion_message=response_text,
+                transformed_generation=extracted_answer,
+            ),
+            expected_output=dataset_sample.data["Answer"],
         )
 
-    def aggregate_results(self, eval_results: List[SingleEvalResult]) -> EvalResult:
-        print("aggregate_results", eval_results)
-        sum_score = sum([result.score_data["score"] for result in eval_results])
+    # def score_sample(self, sample: ProcessedDictSample) -> SingleEvalResult:
+    #     postprocessed_output = sample.postprocessed["postprocessed"]
+    #     expected_answer = sample.data["Answer"]
 
-        return EvalResult(metrics={"score": str(sum_score / len(eval_results))})
+    #     extracted_answer = None
+    #     for answer_regex in MULTILINGUAL_ANSWER_REGEXES:
+    #         regex = MULTILINGUAL_ANSWER_PATTERN_TEMPLATE.format(answer_regex)
+    #         match = re.search(regex, postprocessed_output)
+    #         if match:
+    #             extracted_answer = normalize_extracted_answer(match.group(1))
+    #             break
+
+    #     score = 1.0 if extracted_answer and extracted_answer == expected_answer else 0.0
+
+    #     return SingleEvalResult(
+    #         score_data={
+    #             "score": score,
+    #         },
+    #     )
+
+    # def aggregate_results(self, eval_results: List[SingleEvalResult]) -> EvalResult:
+    #     print("aggregate_results", eval_results)
+    #     sum_score = sum([result.score_data["score"] for result in eval_results])
+
+    #     return EvalResult(metrics={"score": str(sum_score / len(eval_results))})

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

@@ -0,0 +1,5 @@
+# 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.

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

@@ -0,0 +1,78 @@
+# 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.
+import random
+
+from llama_stack.apis.evals.evals import BaseScorer, EvalResult, SingleEvalResult
+from llama_stack.apis.dataset.dataset 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()})
+
+    def aggregate_results(self, eval_results: List[SingleEvalResult]) -> EvalResult:
+        avg_random = sum(
+            [result.score_data["random"] for result in eval_results]
+        ) / len(eval_results)
+        max_random = max([result.score_data["random"] for result in eval_results])
+        return EvalResult(
+            metrics={
+                "avg_random": avg_random,
+                "max_random": max_random,
+            }
+        )
+
+
+class AccuracyScorer(BaseScorer[ScorerInputSample]):
+    def score_sample(self, scorer_input_sample: ScorerInputSample) -> SingleEvalResult:
+        extracted_answer = scorer_input_sample.generation_output.transformed_generation
+        expected_answer = scorer_input_sample.expected_output
+
+        accuracy = (
+            1.0 if extracted_answer and extracted_answer == expected_answer else 0.0
+        )
+
+        return SingleEvalResult(score_data={"accuracy": accuracy})
+
+    def aggregate_results(self, eval_results: List[SingleEvalResult]) -> EvalResult:
+        num_correct = sum([result.score_data["accuracy"] for result in eval_results])
+        num_total = len(eval_results)
+
+        return EvalResult(
+            metrics={
+                "avg_accuracy": num_correct / num_total,
+                "num_correct": num_correct,
+                "num_total": num_total,
+            }
+        )

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

@@ -0,0 +1,39 @@
+# 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.distribution.registry.datasets.dataset_registry import DatasetRegistry
+
+from llama_stack.apis.evals import *  # noqa: F403
+
+
+class RunEvalTask(BaseTask):
+    """
+    RunEvalTask for LlamaStack
+    """
+
+    def __init__(
+        self,
+        eval_task_config,
+        generator_processor: Optional[BaseGeneratorProcessor] = None,
+        generator: Optional[BaseGenerator] = None,
+        scorer: Optional[BaseScorer] = None,
+        *args,
+        **kwargs,
+    ) -> None:
+        super().__init__(
+            generator_processor=generator_processor,
+            generator=generator,
+            scorer=scorer,
+            *args,
+            **kwargs,
+        )
+        self.eval_task_config = eval_task_config
+        self.dataset = DatasetRegistry.get_dataset(
+            eval_task_config.dataset_config.dataset_name
+        )
+
+    def run(self, *args, **kwargs) -> EvalResult:
+        print(f"Running eval task on {self.dataset}")
+        return EvalResult()