fix tests after registration migration & rename meta-reference -> basic / llm_as_judge provider (#424)

* rename meta-reference -> basic

* config rename

* impl rename

* rename llm_as_judge, fix test

* util

* rebase

* naming fix

llama_stack/providers/inline/scoring/basic/__init__.py

@@ -0,0 +1,25 @@
+# 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 Dict
+
+from llama_stack.distribution.datatypes import Api, ProviderSpec
+
+from .config import BasicScoringConfig
+
+
+async def get_provider_impl(
+    config: BasicScoringConfig,
+    deps: Dict[Api, ProviderSpec],
+):
+    from .scoring import BasicScoringImpl
+
+    impl = BasicScoringImpl(
+        config,
+        deps[Api.datasetio],
+        deps[Api.datasets],
+    )
+    await impl.initialize()
+    return impl

llama_stack/providers/inline/scoring/basic/config.py

@@ -0,0 +1,9 @@
+# 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 pydantic import BaseModel
+
+
+class BasicScoringConfig(BaseModel): ...

llama_stack/providers/inline/scoring/basic/scoring.py

@@ -0,0 +1,124 @@
+# 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 List
+
+from llama_models.llama3.api.datatypes import *  # noqa: F403
+from llama_stack.apis.scoring import *  # noqa: F403
+from llama_stack.apis.scoring_functions import *  # noqa: F403
+from llama_stack.apis.common.type_system import *  # noqa: F403
+from llama_stack.apis.datasetio import *  # noqa: F403
+from llama_stack.apis.datasets import *  # noqa: F403
+from llama_stack.providers.datatypes import ScoringFunctionsProtocolPrivate
+
+from .config import BasicScoringConfig
+from .scoring_fn.equality_scoring_fn import EqualityScoringFn
+from .scoring_fn.regex_parser_scoring_fn import RegexParserScoringFn
+from .scoring_fn.subset_of_scoring_fn import SubsetOfScoringFn
+
+FIXED_FNS = [EqualityScoringFn, SubsetOfScoringFn, RegexParserScoringFn]
+
+
+class BasicScoringImpl(Scoring, ScoringFunctionsProtocolPrivate):
+    def __init__(
+        self,
+        config: BasicScoringConfig,
+        datasetio_api: DatasetIO,
+        datasets_api: Datasets,
+    ) -> None:
+        self.config = config
+        self.datasetio_api = datasetio_api
+        self.datasets_api = datasets_api
+        self.scoring_fn_id_impls = {}
+
+    async def initialize(self) -> None:
+        for fn in FIXED_FNS:
+            impl = fn()
+            for fn_defs in impl.get_supported_scoring_fn_defs():
+                self.scoring_fn_id_impls[fn_defs.identifier] = impl
+
+    async def shutdown(self) -> None: ...
+
+    async def list_scoring_functions(self) -> List[ScoringFn]:
+        scoring_fn_defs_list = [
+            fn_def
+            for impl in self.scoring_fn_id_impls.values()
+            for fn_def in impl.get_supported_scoring_fn_defs()
+        ]
+
+        for f in scoring_fn_defs_list:
+            assert f.identifier.startswith(
+                "basic"
+            ), "All basic scoring fn must have identifier prefixed with 'basic'! "
+
+        return scoring_fn_defs_list
+
+    async def register_scoring_function(self, function_def: ScoringFn) -> None:
+        raise NotImplementedError("Register scoring function not implemented yet")
+
+    async def validate_scoring_input_dataset_schema(self, dataset_id: str) -> None:
+        dataset_def = await self.datasets_api.get_dataset(dataset_id=dataset_id)
+        if not dataset_def.schema or len(dataset_def.schema) == 0:
+            raise ValueError(
+                f"Dataset {dataset_id} does not have a schema defined. Please define a schema for the dataset."
+            )
+
+        for required_column in ["generated_answer", "expected_answer", "input_query"]:
+            if required_column not in dataset_def.schema:
+                raise ValueError(
+                    f"Dataset {dataset_id} does not have a '{required_column}' column."
+                )
+            if dataset_def.schema[required_column].type != "string":
+                raise ValueError(
+                    f"Dataset {dataset_id} does not have a '{required_column}' column of type 'string'."
+                )
+
+    async def score_batch(
+        self,
+        dataset_id: str,
+        scoring_functions: Dict[str, Optional[ScoringFnParams]] = None,
+        save_results_dataset: bool = False,
+    ) -> ScoreBatchResponse:
+        await self.validate_scoring_input_dataset_schema(dataset_id=dataset_id)
+        all_rows = await self.datasetio_api.get_rows_paginated(
+            dataset_id=dataset_id,
+            rows_in_page=-1,
+        )
+        res = await self.score(
+            input_rows=all_rows.rows,
+            scoring_functions=scoring_functions,
+        )
+        if save_results_dataset:
+            # TODO: persist and register dataset on to server for reading
+            # self.datasets_api.register_dataset()
+            raise NotImplementedError("Save results dataset not implemented yet")
+
+        return ScoreBatchResponse(
+            results=res.results,
+        )
+
+    async def score(
+        self,
+        input_rows: List[Dict[str, Any]],
+        scoring_functions: Dict[str, Optional[ScoringFnParams]] = None,
+    ) -> ScoreResponse:
+        res = {}
+        for scoring_fn_id in scoring_functions.keys():
+            if scoring_fn_id not in self.scoring_fn_id_impls:
+                raise ValueError(f"Scoring function {scoring_fn_id} is not supported.")
+            scoring_fn = self.scoring_fn_id_impls[scoring_fn_id]
+            scoring_fn_params = scoring_functions.get(scoring_fn_id, None)
+            score_results = await scoring_fn.score(
+                input_rows, scoring_fn_id, scoring_fn_params
+            )
+            agg_results = await scoring_fn.aggregate(score_results)
+            res[scoring_fn_id] = ScoringResult(
+                score_rows=score_results,
+                aggregated_results=agg_results,
+            )
+
+        return ScoreResponse(
+            results=res,
+        )

llama_stack/providers/inline/scoring/basic/scoring_fn/__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/inline/scoring/basic/scoring_fn/equality_scoring_fn.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 llama_stack.providers.utils.scoring.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 llama_stack.providers.utils.scoring.aggregation_utils import aggregate_accuracy
+
+from .fn_defs.equality import equality
+
+
+class EqualityScoringFn(BaseScoringFn):
+    """
+    A scoring_fn that assigns a score of 1.0 if the input string matches the target string, and 0.0 otherwise.
+    """
+
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+        self.supported_fn_defs_registry = {
+            equality.identifier: equality,
+        }
+
+    async def score_row(
+        self,
+        input_row: Dict[str, Any],
+        scoring_fn_identifier: Optional[str] = "equality",
+        scoring_params: Optional[ScoringFnParams] = None,
+    ) -> ScoringResultRow:
+        assert "expected_answer" in input_row, "Expected answer not found in input row."
+        assert (
+            "generated_answer" in input_row
+        ), "Generated answer not found in input row."
+
+        expected_answer = input_row["expected_answer"]
+        generated_answer = input_row["generated_answer"]
+        score = 1.0 if expected_answer == generated_answer else 0.0
+        return {
+            "score": score,
+        }
+
+    async def aggregate(
+        self, scoring_results: List[ScoringResultRow]
+    ) -> Dict[str, Any]:
+        return aggregate_accuracy(scoring_results)

llama_stack/providers/inline/scoring/basic/scoring_fn/fn_defs/__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/inline/scoring/basic/scoring_fn/fn_defs/equality.py

@@ -0,0 +1,18 @@
+# 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 ScoringFn
+
+
+equality = ScoringFn(
+    identifier="basic::equality",
+    description="Returns 1.0 if the input is equal to the target, 0.0 otherwise.",
+    params=None,
+    provider_id="basic",
+    provider_resource_id="equality",
+    return_type=NumberType(),
+)

llama_stack/providers/inline/scoring/basic/scoring_fn/fn_defs/regex_parser_multiple_choice_answer.py

@@ -0,0 +1,71 @@
+# 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.scoring_functions import *  # noqa: F401, F403
+from llama_stack.apis.scoring import *  # noqa: F401, F403
+from llama_stack.apis.common.type_system import NumberType
+
+MULTILINGUAL_ANSWER_REGEXES = [
+    r"Answer\s*:",
+    r"Answer\s*:​​​​​​",  # Korean invisible character
+    r"উত্তর\s*:",
+    r"उत्तर\s*:",
+    r"উত্তরঃ",
+    r"উত্তর\s*:",
+    r"Antwort\s*:",
+    r"답변\s*:",
+    r"정답\s*:",
+    r"답\s*:",
+    r"答案\s*：",
+    r"答案\s*:",
+    r"答\s*：",
+    r"答\s*:",
+    r"答复\s*：",
+    r"答曰\s*：",
+    r"الإجابة:",
+    r"الجواب:",
+    r"إجابة:",
+    r"الإجابة النهائية:",
+    r"الإجابة الصحيحة:",
+    r"الإجابة الصحيحة هي:",
+    r"الإجابة هي:",
+    r"Respuesta\s*:",
+    r"Risposta\s*:",
+    r"答え\s*:",
+    r"答え\s*：",
+    r"回答\s*:",
+    r"回答\s*：",
+    r"解答\s*:",
+    r"Jawaban\s*:",
+    r"Réponse\s*:",
+    r"Resposta\s*:",
+    r"Jibu\s*:",
+    r"Idahun\s*:",
+    r"Ìdáhùn\s*:",
+    r"Idáhùn\s*:",
+    r"Àmọ̀nà\s*:",
+    r"Àdáhùn\s*:",
+    r"Ànúgọ\s*:",
+    r"Àṣàyàn\s*:",
+]
+
+MULTILINGUAL_ANSWER_PATTERN_TEMPLATE = (
+    r"(?i){}\s*([A-D]|[أ-د]|[অ]|[ব]|[ড]|[ঢ]|[Ａ]|[Ｂ]|[Ｃ]|[Ｄ])"
+)
+
+regex_parser_multiple_choice_answer = ScoringFn(
+    identifier="basic::regex_parser_multiple_choice_answer",
+    description="Extract answer from response matching Answer: [the_answer_letter], and compare with expected result",
+    return_type=NumberType(),
+    provider_id="basic",
+    provider_resource_id="regex-parser-multiple-choice-answer",
+    params=RegexParserScoringFnParams(
+        parsing_regexes=[
+            MULTILINGUAL_ANSWER_PATTERN_TEMPLATE.format(x)
+            for x in MULTILINGUAL_ANSWER_REGEXES
+        ],
+    ),
+)

llama_stack/providers/inline/scoring/basic/scoring_fn/fn_defs/subset_of.py

@@ -0,0 +1,17 @@
+# 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 ScoringFn
+
+
+subset_of = ScoringFn(
+    identifier="basic::subset_of",
+    description="Returns 1.0 if the expected is included in generated, 0.0 otherwise.",
+    return_type=NumberType(),
+    provider_id="basic",
+    provider_resource_id="subset-of",
+)

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

@@ -0,0 +1,67 @@
+# 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 re
+
+from llama_stack.providers.utils.scoring.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 llama_stack.providers.utils.scoring.aggregation_utils import aggregate_accuracy
+
+from .fn_defs.regex_parser_multiple_choice_answer import (
+    regex_parser_multiple_choice_answer,
+)
+
+
+class RegexParserScoringFn(BaseScoringFn):
+    """
+    A scoring_fn that parses answer from generated response according to context and check match with expected_answer.
+    """
+
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+        self.supported_fn_defs_registry = {
+            regex_parser_multiple_choice_answer.identifier: regex_parser_multiple_choice_answer,
+        }
+
+    async def score_row(
+        self,
+        input_row: Dict[str, Any],
+        scoring_fn_identifier: Optional[str] = None,
+        scoring_params: Optional[ScoringFnParams] = None,
+    ) -> ScoringResultRow:
+        assert (
+            scoring_fn_identifier is not None
+        ), "Scoring function identifier not found."
+        fn_def = self.supported_fn_defs_registry[scoring_fn_identifier]
+        if scoring_params is not None:
+            fn_def.params = scoring_params
+
+        assert (
+            fn_def.params is not None
+            and fn_def.params.type == ScoringFnParamsType.regex_parser.value
+        ), f"RegexParserScoringFnParams not found for {fn_def}."
+
+        expected_answer = input_row["expected_answer"]
+        generated_answer = input_row["generated_answer"]
+
+        # parse answer according to regex
+        parsed_answer = None
+        for regex in fn_def.params.parsing_regexes:
+            match = re.search(regex, generated_answer)
+            if match:
+                parsed_answer = match.group(1)
+                break
+
+        score = 1.0 if parsed_answer and parsed_answer == expected_answer else 0.0
+        return {
+            "score": score,
+        }
+
+    async def aggregate(
+        self, scoring_results: List[ScoringResultRow]
+    ) -> Dict[str, Any]:
+        return aggregate_accuracy(scoring_results)

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

@@ -0,0 +1,43 @@
+# 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.providers.utils.scoring.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 llama_stack.providers.utils.scoring.aggregation_utils import aggregate_accuracy
+
+from .fn_defs.subset_of import subset_of
+
+
+class SubsetOfScoringFn(BaseScoringFn):
+    """
+    A scoring_fn that assigns a score of 1.0 if the expected string is included in the generated string, and 0.0 otherwise.
+    """
+
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+        self.supported_fn_defs_registry = {
+            subset_of.identifier: subset_of,
+        }
+
+    async def score_row(
+        self,
+        input_row: Dict[str, Any],
+        scoring_fn_identifier: Optional[str] = "subset_of",
+        scoring_params: Optional[ScoringFnParams] = None,
+    ) -> ScoringResultRow:
+        expected_answer = input_row["expected_answer"]
+        generated_answer = input_row["generated_answer"]
+        score = 1.0 if expected_answer in generated_answer else 0.0
+        return {
+            "score": score,
+        }
+
+    async def aggregate(
+        self, scoring_results: List[ScoringResultRow]
+    ) -> Dict[str, Any]:
+        return aggregate_accuracy(scoring_results)