evals new rebase

llama_stack/distribution/registry/tasks/__init__.py

@@ -0,0 +1,13 @@
+# 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.
+# TODO: make these import config based
+from llama_stack.providers.impls.meta_reference.evals.tasks.mmlu_task import MMLUTask
+from .task_registry import TaskRegistry
+
+TaskRegistry.register(
+    "mmlu",
+    MMLUTask,
+)

llama_stack/distribution/registry/tasks/task.py

@@ -0,0 +1,49 @@
+# 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 abc import ABC, abstractmethod
+
+
+class BaseTask(ABC):
+    """
+    A task represents a single evaluation benchmark, including it's dataset, preprocessing, postprocessing and scoring methods.
+    Base class for all evaluation tasks. Each task needs to implement the following methods:
+    - F1: preprocess_sample(self)
+    - F2: postprocess_sample(self)
+    - F3: score_sample(self)
+    """
+
+    def __init__(self, dataset, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._name = self.__class__.__name__
+        self.dataset = dataset
+
+    @abstractmethod
+    def preprocess_sample(self, sample):
+        raise NotImplementedError()
+
+    @abstractmethod
+    def postprocess_sample(self, sample):
+        raise NotImplementedError()
+
+    @abstractmethod
+    def score_sample(self, sample, ground_truth):
+        raise NotImplementedError()
+
+    @abstractmethod
+    def aggregate_results(self, eval_results):
+        raise NotImplementedError()
+
+    def preprocess(self):
+        return [self.preprocess_sample(sample) for sample in self.dataset]
+
+    def postprocess(self, generation):
+        return [self.postprocess_sample(sample) for sample in generation]
+
+    def score(self, postprocessed):
+        return [
+            self.score_sample(sample, ground_truth)
+            for sample, ground_truth in zip(postprocessed, self.dataset)
+        ]

llama_stack/distribution/registry/tasks/task_registry.py

@@ -0,0 +1,32 @@
+# 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 typing import AbstractSet, Dict
+
+from .task import BaseTask
+
+
+class TaskRegistry:
+    _REGISTRY: Dict[str, BaseTask] = {}
+
+    @staticmethod
+    def names() -> AbstractSet[str]:
+        return TaskRegistry._REGISTRY.keys()
+
+    @staticmethod
+    def register(name: str, task: BaseTask) -> None:
+        if name in TaskRegistry._REGISTRY:
+            raise ValueError(f"Task {name} already exists.")
+        TaskRegistry._REGISTRY[name] = task
+
+    @staticmethod
+    def get_task(name: str) -> BaseTask:
+        if name not in TaskRegistry._REGISTRY:
+            raise ValueError(f"Task {name} not found.")
+        return TaskRegistry._REGISTRY[name]
+
+    @staticmethod
+    def reset() -> None:
+        TaskRegistry._REGISTRY = {}