Folder restructure for evals/datasets/scoring (#419)

* rename evals related stuff

* fix datasetio

* fix scoring test

* localfs -> LocalFS

* refactor scoring

* refactor scoring

* remove 8b_correctness scoring_fn from tests

* tests w/ eval params

* scoring fn braintrust fixture

* import

llama_stack/providers/inline/datasetio/localfs/__init__.py

@@ -4,15 +4,15 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
-from .config import MetaReferenceDatasetIOConfig
+from .config import LocalFSDatasetIOConfig
 
 
 async def get_provider_impl(
-    config: MetaReferenceDatasetIOConfig,
+    config: LocalFSDatasetIOConfig,
     _deps,
 ):
-    from .datasetio import MetaReferenceDatasetIOImpl
+    from .datasetio import LocalFSDatasetIOImpl
 
-    impl = MetaReferenceDatasetIOImpl(config)
+    impl = LocalFSDatasetIOImpl(config)
     await impl.initialize()
     return impl

llama_stack/providers/inline/datasetio/localfs/config.py

@@ -6,4 +6,4 @@
 from llama_stack.apis.datasetio import *  # noqa: F401, F403
 
 
-class MetaReferenceDatasetIOConfig(BaseModel): ...
+class LocalFSDatasetIOConfig(BaseModel): ...

llama_stack/providers/inline/datasetio/localfs/datasetio.py

@@ -15,7 +15,7 @@ from dataclasses import dataclass
 from llama_stack.providers.datatypes import DatasetsProtocolPrivate
 from llama_stack.providers.utils.datasetio.url_utils import get_dataframe_from_url
 
-from .config import MetaReferenceDatasetIOConfig
+from .config import LocalFSDatasetIOConfig
 
 
 class BaseDataset(ABC):
@@ -77,8 +77,8 @@ class PandasDataframeDataset(BaseDataset):
         self.df = self._validate_dataset_schema(df)
 
 
-class MetaReferenceDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
+class LocalFSDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
-    def __init__(self, config: MetaReferenceDatasetIOConfig) -> None:
+    def __init__(self, config: LocalFSDatasetIOConfig) -> None:
         self.config = config
         # local registry for keeping track of datasets within the provider
         self.dataset_infos = {}

llama_stack/providers/inline/eval/meta_reference/eval.py

@@ -9,14 +9,13 @@ from llama_models.llama3.api.datatypes import *  # noqa: F403
 from .....apis.common.job_types import Job
 from .....apis.eval.eval import Eval, EvalTaskConfig, EvaluateResponse, JobStatus
 from llama_stack.apis.common.type_system import *  # noqa: F403
-from tqdm import tqdm
-
 from llama_stack.apis.datasetio import DatasetIO
 from llama_stack.apis.datasets import Datasets
 from llama_stack.apis.eval_tasks import EvalTaskDef
 from llama_stack.apis.inference import Inference
 from llama_stack.apis.scoring import Scoring
 from llama_stack.providers.datatypes import EvalTasksProtocolPrivate
+from tqdm import tqdm
 
 from .config import MetaReferenceEvalConfig
 

llama_stack/providers/inline/scoring/braintrust/braintrust.py

@@ -16,9 +16,8 @@ from llama_stack.apis.datasets import *  # noqa: F403
 from autoevals.llm import Factuality
 from autoevals.ragas import AnswerCorrectness
 from llama_stack.providers.datatypes import ScoringFunctionsProtocolPrivate
-from llama_stack.providers.inline.meta_reference.scoring.scoring_fn.common import (
+
-    aggregate_average,
+from llama_stack.providers.utils.scoring.aggregation_utils import aggregate_average
-)
 
 from .config import BraintrustScoringConfig
 from .scoring_fn.fn_defs.answer_correctness import answer_correctness_fn_def

llama_stack/providers/inline/scoring/meta_reference/scoring_fn/equality_scoring_fn.py

@@ -4,20 +4,14 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
-from llama_stack.providers.inline.meta_reference.scoring.scoring_fn.base_scoring_fn import (
+from .base_scoring_fn import BaseScoringFn
-    BaseScoringFn,
-)
 from llama_stack.apis.scoring_functions import *  # noqa: F401, F403
 from llama_stack.apis.scoring import *  # noqa: F401, F403
 from llama_stack.apis.common.type_system import *  # noqa: F403
 
-from llama_stack.providers.inline.meta_reference.scoring.scoring_fn.common import (
+from llama_stack.providers.utils.scoring.aggregation_utils import aggregate_accuracy
-    aggregate_accuracy,
-)
 
-from llama_stack.providers.inline.meta_reference.scoring.scoring_fn.fn_defs.equality import (
+from .fn_defs.equality import equality
-    equality,
-)
 
 
 class EqualityScoringFn(BaseScoringFn):

llama_stack/providers/inline/scoring/meta_reference/scoring_fn/fn_defs/llm_as_judge_base.py

@@ -0,0 +1,15 @@
+# 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.common.type_system import NumberType
+from llama_stack.apis.scoring_functions import ScoringFnDef
+
+
+llm_as_judge_base = ScoringFnDef(
+    identifier="meta-reference::llm_as_judge_base",
+    description="Llm As Judge Scoring Function",
+    return_type=NumberType(),
+)

llama_stack/providers/inline/scoring/meta_reference/scoring_fn/llm_as_judge_scoring_fn.py

@@ -4,20 +4,16 @@
 # 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.inference.inference import Inference
-from llama_stack.providers.inline.meta_reference.scoring.scoring_fn.base_scoring_fn import (
+
-    BaseScoringFn,
+from .base_scoring_fn import BaseScoringFn
-)
 from llama_stack.apis.scoring_functions import *  # noqa: F401, F403
 from llama_stack.apis.scoring import *  # noqa: F401, F403
 from llama_stack.apis.common.type_system import *  # noqa: F403
 import re
 
-from llama_stack.providers.inline.meta_reference.scoring.scoring_fn.common import (
+from llama_stack.providers.utils.scoring.aggregation_utils import aggregate_average
-    aggregate_average,
+
-)
+from .fn_defs.llm_as_judge_base import llm_as_judge_base
-from llama_stack.providers.inline.meta_reference.scoring.scoring_fn.fn_defs.llm_as_judge_8b_correctness import (
-    llm_as_judge_8b_correctness,
-)
 
 
 class LlmAsJudgeScoringFn(BaseScoringFn):
@@ -29,7 +25,7 @@ class LlmAsJudgeScoringFn(BaseScoringFn):
         super().__init__(*arg, **kwargs)
         self.inference_api = inference_api
         self.supported_fn_defs_registry = {
-            llm_as_judge_8b_correctness.identifier: llm_as_judge_8b_correctness,
+            llm_as_judge_base.identifier: llm_as_judge_base,
         }
 
     async def score_row(

llama_stack/providers/inline/scoring/meta_reference/scoring_fn/regex_parser_scoring_fn.py

@@ -9,7 +9,7 @@ from .base_scoring_fn import BaseScoringFn
 from llama_stack.apis.scoring_functions import *  # noqa: F401, F403
 from llama_stack.apis.scoring import *  # noqa: F401, F403
 from llama_stack.apis.common.type_system import *  # noqa: F403
-from .common import aggregate_accuracy
+from llama_stack.providers.utils.scoring.aggregation_utils import aggregate_accuracy
 
 from .fn_defs.regex_parser_multiple_choice_answer import (
     regex_parser_multiple_choice_answer,

llama_stack/providers/inline/scoring/meta_reference/scoring_fn/subset_of_scoring_fn.py

@@ -4,19 +4,13 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
-from llama_stack.providers.inline.meta_reference.scoring.scoring_fn.base_scoring_fn import (
+from .base_scoring_fn import BaseScoringFn
-    BaseScoringFn,
-)
 from llama_stack.apis.scoring_functions import *  # noqa: F401, F403
 from llama_stack.apis.scoring import *  # noqa: F401, F403
 from llama_stack.apis.common.type_system import *  # noqa: F403
-from llama_stack.providers.inline.meta_reference.scoring.scoring_fn.common import (
+from llama_stack.providers.utils.scoring.aggregation_utils import aggregate_accuracy
-    aggregate_accuracy,
-)
 
-from llama_stack.providers.inline.meta_reference.scoring.scoring_fn.fn_defs.subset_of import (
+from .fn_defs.subset_of import subset_of
-    subset_of,
-)
 
 
 class SubsetOfScoringFn(BaseScoringFn):

llama_stack/providers/registry/datasetio.py

@@ -13,10 +13,10 @@ def available_providers() -> List[ProviderSpec]:
     return [
         InlineProviderSpec(
             api=Api.datasetio,
-            provider_type="meta-reference",
+            provider_type="localfs",
             pip_packages=["pandas"],
-            module="llama_stack.providers.inline.meta_reference.datasetio",
+            module="llama_stack.providers.inline.datasetio.localfs",
-            config_class="llama_stack.providers.inline.meta_reference.datasetio.MetaReferenceDatasetIOConfig",
+            config_class="llama_stack.providers.inline.datasetio.localfs.LocalFSDatasetIOConfig",
             api_dependencies=[],
         ),
         remote_provider_spec(

llama_stack/providers/registry/eval.py

@@ -15,8 +15,8 @@ def available_providers() -> List[ProviderSpec]:
             api=Api.eval,
             provider_type="meta-reference",
             pip_packages=[],
-            module="llama_stack.providers.inline.meta_reference.eval",
+            module="llama_stack.providers.inline.eval.meta_reference",
-            config_class="llama_stack.providers.inline.meta_reference.eval.MetaReferenceEvalConfig",
+            config_class="llama_stack.providers.inline.eval.meta_reference.MetaReferenceEvalConfig",
             api_dependencies=[
                 Api.datasetio,
                 Api.datasets,

llama_stack/providers/registry/scoring.py

@@ -15,8 +15,8 @@ def available_providers() -> List[ProviderSpec]:
             api=Api.scoring,
             provider_type="meta-reference",
             pip_packages=[],
-            module="llama_stack.providers.inline.meta_reference.scoring",
+            module="llama_stack.providers.inline.scoring.meta_reference",
-            config_class="llama_stack.providers.inline.meta_reference.scoring.MetaReferenceScoringConfig",
+            config_class="llama_stack.providers.inline.scoring.meta_reference.MetaReferenceScoringConfig",
             api_dependencies=[
                 Api.datasetio,
                 Api.datasets,
@@ -27,8 +27,8 @@ def available_providers() -> List[ProviderSpec]:
             api=Api.scoring,
             provider_type="braintrust",
             pip_packages=["autoevals", "openai"],
-            module="llama_stack.providers.inline.braintrust.scoring",
+            module="llama_stack.providers.inline.scoring.braintrust",
-            config_class="llama_stack.providers.inline.braintrust.scoring.BraintrustScoringConfig",
+            config_class="llama_stack.providers.inline.scoring.braintrust.BraintrustScoringConfig",
             api_dependencies=[
                 Api.datasetio,
                 Api.datasets,

llama_stack/providers/tests/datasetio/fixtures.py

@@ -19,12 +19,12 @@ def datasetio_remote() -> ProviderFixture:
 
 
 @pytest.fixture(scope="session")
-def datasetio_meta_reference() -> ProviderFixture:
+def datasetio_localfs() -> ProviderFixture:
     return ProviderFixture(
         providers=[
             Provider(
-                provider_id="meta-reference",
+                provider_id="localfs",
-                provider_type="meta-reference",
+                provider_type="localfs",
                 config={},
             )
         ],
@@ -44,7 +44,7 @@ def datasetio_huggingface() -> ProviderFixture:
     )
 
 
-DATASETIO_FIXTURES = ["meta_reference", "remote", "huggingface"]
+DATASETIO_FIXTURES = ["localfs", "remote", "huggingface"]
 
 
 @pytest_asyncio.fixture(scope="session")

llama_stack/providers/tests/eval/constants.py

@@ -4,10 +4,6 @@
 # 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.scoring_functions import *  # noqa: F401, F403
-from llama_stack.apis.scoring import *  # noqa: F401, F403
-from llama_stack.apis.common.type_system import NumberType
-
 JUDGE_PROMPT = """
 You will be given a question, a expected_answer, and a system_answer.
 Your task is to provide a 'total rating' scoring how well the system_answer answers compared with ground truth in expected_answer in terms of factual correctness to the question.
@@ -22,18 +18,3 @@ System Answer: {generated_answer}
 Feedback:::
 Total rating:
 """
-
-llm_as_judge_8b_correctness = ScoringFnDef(
-    identifier="meta-reference::llm_as_judge_8b_correctness",
-    description="Llm As Judge Scoring Function",
-    return_type=NumberType(),
-    params=LLMAsJudgeScoringFnParams(
-        prompt_template=JUDGE_PROMPT,
-        judge_model="Llama3.1-8B-Instruct",
-        judge_score_regexes=[
-            r"Total rating: (\d+)",
-            r"rating: (\d+)",
-            r"Rating: (\d+)",
-        ],
-    ),
-)

llama_stack/providers/tests/eval/test_eval.py

@@ -19,9 +19,10 @@ from llama_stack.apis.eval.eval import (
     EvalTaskDefWithProvider,
     ModelCandidate,
 )
+from llama_stack.apis.scoring_functions import LLMAsJudgeScoringFnParams
 from llama_stack.distribution.datatypes import Api
 from llama_stack.providers.tests.datasetio.test_datasetio import register_dataset
-
+from .constants import JUDGE_PROMPT
 
 # How to run this test:
 #
@@ -65,7 +66,7 @@ class Testeval:
         assert len(rows.rows) == 3
 
         scoring_functions = [
-            "meta-reference::llm_as_judge_8b_correctness",
+            "meta-reference::llm_as_judge_base",
             "meta-reference::equality",
         ]
         task_id = "meta-reference::app_eval"
@@ -85,11 +86,22 @@ class Testeval:
                     model="Llama3.2-3B-Instruct",
                     sampling_params=SamplingParams(),
                 ),
+                scoring_params={
+                    "meta-reference::llm_as_judge_base": LLMAsJudgeScoringFnParams(
+                        judge_model="Llama3.1-8B-Instruct",
+                        prompt_template=JUDGE_PROMPT,
+                        judge_score_regexes=[
+                            r"Total rating: (\d+)",
+                            r"rating: (\d+)",
+                            r"Rating: (\d+)",
+                        ],
+                    )
+                },
             ),
         )
         assert len(response.generations) == 3
-        assert "meta-reference::llm_as_judge_8b_correctness" in response.scores
         assert "meta-reference::equality" in response.scores
+        assert "meta-reference::llm_as_judge_base" in response.scores
 
     @pytest.mark.asyncio
     async def test_eval_run_eval(self, eval_stack):
@@ -109,7 +121,6 @@ class Testeval:
         )
 
         scoring_functions = [
-            "meta-reference::llm_as_judge_8b_correctness",
             "meta-reference::subset_of",
         ]
 
@@ -138,7 +149,6 @@ class Testeval:
         assert eval_response is not None
         assert len(eval_response.generations) == 5
         assert "meta-reference::subset_of" in eval_response.scores
-        assert "meta-reference::llm_as_judge_8b_correctness" in eval_response.scores
 
     @pytest.mark.asyncio
     async def test_eval_run_benchmark_eval(self, eval_stack):

llama_stack/providers/tests/scoring/conftest.py

@@ -16,7 +16,7 @@ DEFAULT_PROVIDER_COMBINATIONS = [
     pytest.param(
         {
             "scoring": "meta_reference",
-            "datasetio": "meta_reference",
+            "datasetio": "localfs",
             "inference": "fireworks",
         },
         id="meta_reference_scoring_fireworks_inference",
@@ -25,12 +25,21 @@ DEFAULT_PROVIDER_COMBINATIONS = [
     pytest.param(
         {
             "scoring": "meta_reference",
-            "datasetio": "meta_reference",
+            "datasetio": "localfs",
             "inference": "together",
         },
         id="meta_reference_scoring_together_inference",
         marks=pytest.mark.meta_reference_scoring_together_inference,
     ),
+    pytest.param(
+        {
+            "scoring": "braintrust",
+            "datasetio": "localfs",
+            "inference": "together",
+        },
+        id="braintrust_scoring_together_inference",
+        marks=pytest.mark.braintrust_scoring_together_inference,
+    ),
 ]
 
 
@@ -38,6 +47,7 @@ def pytest_configure(config):
     for fixture_name in [
         "meta_reference_scoring_fireworks_inference",
         "meta_reference_scoring_together_inference",
+        "braintrust_scoring_together_inference",
     ]:
         config.addinivalue_line(
             "markers",

llama_stack/providers/tests/scoring/fixtures.py

@@ -31,7 +31,20 @@ def scoring_meta_reference() -> ProviderFixture:
     )
 
 
-SCORING_FIXTURES = ["meta_reference", "remote"]
+@pytest.fixture(scope="session")
+def scoring_braintrust() -> ProviderFixture:
+    return ProviderFixture(
+        providers=[
+            Provider(
+                provider_id="braintrust",
+                provider_type="braintrust",
+                config={},
+            )
+        ],
+    )
+
+
+SCORING_FIXTURES = ["meta_reference", "remote", "braintrust"]
 
 
 @pytest_asyncio.fixture(scope="session")
@@ -52,9 +65,4 @@ async def scoring_stack(request):
         provider_data,
     )
 
-    return (
+    return impls
-        impls[Api.scoring],
-        impls[Api.scoring_functions],
-        impls[Api.datasetio],
-        impls[Api.datasets],
-    )

llama_stack/providers/tests/scoring/test_scoring.py

@@ -8,7 +8,7 @@
 import pytest
 
 from llama_stack.apis.scoring_functions import *  # noqa: F403
-
+from llama_stack.distribution.datatypes import Api
 from llama_stack.providers.tests.datasetio.test_datasetio import register_dataset
 
 # How to run this test:
@@ -23,20 +23,36 @@ class TestScoring:
     async def test_scoring_functions_list(self, scoring_stack):
         # NOTE: this needs you to ensure that you are starting from a clean state
         # but so far we don't have an unregister API unfortunately, so be careful
-        _, scoring_functions_impl, _, _ = scoring_stack
+        scoring_functions_impl = scoring_stack[Api.scoring_functions]
         response = await scoring_functions_impl.list_scoring_functions()
         assert isinstance(response, list)
         assert len(response) > 0
 
     @pytest.mark.asyncio
     async def test_scoring_score(self, scoring_stack):
-        scoring_impl, scoring_functions_impl, datasetio_impl, datasets_impl = (
+        (
-            scoring_stack
+            scoring_impl,
+            scoring_functions_impl,
+            datasetio_impl,
+            datasets_impl,
+            models_impl,
+        ) = (
+            scoring_stack[Api.scoring],
+            scoring_stack[Api.scoring_functions],
+            scoring_stack[Api.datasetio],
+            scoring_stack[Api.datasets],
+            scoring_stack[Api.models],
         )
         await register_dataset(datasets_impl)
         response = await datasets_impl.list_datasets()
         assert len(response) == 1
 
+        for model_id in ["Llama3.2-3B-Instruct", "Llama3.1-8B-Instruct"]:
+            await models_impl.register_model(
+                model_id=model_id,
+                provider_id="",
+            )
+
         # scoring individual rows
         rows = await datasetio_impl.get_rows_paginated(
             dataset_id="test_dataset",
@@ -44,10 +60,11 @@ class TestScoring:
         )
         assert len(rows.rows) == 3
 
+        scoring_fns_list = await scoring_functions_impl.list_scoring_functions()
         scoring_functions = {
-            "meta-reference::llm_as_judge_8b_correctness": None,
+            scoring_fns_list[0].identifier: None,
-            "meta-reference::equality": None,
         }
+
         response = await scoring_impl.score(
             input_rows=rows.rows,
             scoring_functions=scoring_functions,
@@ -69,13 +86,34 @@ class TestScoring:
 
     @pytest.mark.asyncio
     async def test_scoring_score_with_params(self, scoring_stack):
-        scoring_impl, scoring_functions_impl, datasetio_impl, datasets_impl = (
+        (
-            scoring_stack
+            scoring_impl,
+            scoring_functions_impl,
+            datasetio_impl,
+            datasets_impl,
+            models_impl,
+        ) = (
+            scoring_stack[Api.scoring],
+            scoring_stack[Api.scoring_functions],
+            scoring_stack[Api.datasetio],
+            scoring_stack[Api.datasets],
+            scoring_stack[Api.models],
         )
         await register_dataset(datasets_impl)
         response = await datasets_impl.list_datasets()
         assert len(response) == 1
 
+        for model_id in ["Llama3.1-405B-Instruct"]:
+            await models_impl.register_model(
+                model_id=model_id,
+                provider_id="",
+            )
+
+        scoring_fns_list = await scoring_functions_impl.list_scoring_functions()
+        provider_id = scoring_fns_list[0].provider_id
+        if provider_id == "braintrust":
+            pytest.skip("Braintrust provider does not support scoring with params")
+
         # scoring individual rows
         rows = await datasetio_impl.get_rows_paginated(
             dataset_id="test_dataset",
@@ -84,7 +122,7 @@ class TestScoring:
         assert len(rows.rows) == 3
 
         scoring_functions = {
-            "meta-reference::llm_as_judge_8b_correctness": LLMAsJudgeScoringFnParams(
+            "meta-reference::llm_as_judge_base": LLMAsJudgeScoringFnParams(
                 judge_model="Llama3.1-405B-Instruct",
                 prompt_template="Output a number response in the following format: Score: <number>, where <number> is the number between 0 and 9.",
                 judge_score_regexes=[r"Score: (\d+)"],

llama_stack/providers/utils/scoring/aggregation_utils.py

@@ -3,13 +3,10 @@
 #
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
-from pathlib import Path
 from typing import Any, Dict, List
 
 from llama_stack.apis.scoring import ScoringResultRow
 
-FN_DEFS_PATH = Path(__file__).parent / "fn_defs"
-
 
 def aggregate_accuracy(scoring_results: List[ScoringResultRow]) -> Dict[str, Any]:
     num_correct = sum(result["score"] for result in scoring_results)