evals new rebase

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

@@ -0,0 +1,19 @@
+# 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 .config import MetaReferenceEvalsImplConfig  # noqa
+from llama_stack.apis.inference import *  # noqa: F403
+from llama_stack.distribution.datatypes import Api, ProviderSpec
+
+
+async def get_provider_impl(
+    config: MetaReferenceEvalsImplConfig, deps: Dict[Api, ProviderSpec]
+):
+    from .evals import MetaReferenceEvalsImpl
+
+    impl = MetaReferenceEvalsImpl(config, deps[Api.inference])
+    await impl.initialize()
+    return impl

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

@@ -0,0 +1,10 @@
+# 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 MetaReferenceEvalsImplConfig(BaseModel): ...

llama_stack/providers/impls/meta_reference/evals/evals.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.inference import *  # noqa: F403
+from llama_stack.apis.evals import *  # noqa: F403
+from termcolor import cprint
+
+from llama_stack.distribution.registry.datasets.dataset_registry import DatasetRegistry
+
+from llama_stack.distribution.registry.tasks.task_registry import TaskRegistry
+
+from .config import MetaReferenceEvalsImplConfig
+
+
+class MetaReferenceEvalsImpl(Evals):
+    def __init__(self, config: MetaReferenceEvalsImplConfig, inference_api: Inference):
+        self.inference_api = inference_api
+
+    async def initialize(self) -> None:
+        pass
+
+    async def shutdown(self) -> None:
+        pass
+
+    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)(dataset)
+        x1 = task_impl.preprocess()
+
+        # TODO: replace w/ batch inference & async return eval job
+        generation_outputs = []
+        if eval_task_config is None:
+            eval_task_config = EvaluateTaskConfig(n_samples=len(x1))
+        if eval_task_config.n_samples is None or eval_task_config.n_samples > len(x1):
+            eval_task_config.n_samples = len(x1)
+
+        print(
+            f"Eval generation start, generate on {eval_task_config.n_samples} samples"
+        )
+
+        for msg in x1[: eval_task_config.n_samples]:
+            print("generation for msg: ", msg)
+            response = await self.inference_api.chat_completion(
+                model=model,
+                messages=[msg],
+                stream=False,
+            )
+            generation_outputs.append(response.completion_message.content)
+
+        x2 = task_impl.postprocess(generation_outputs)
+        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/__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/tasks/mmlu_task.py

@@ -0,0 +1,150 @@
+# 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.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: $LETTER' (without quotes) where LETTER is one of ABCD.
+
+{Question}
+
+A) {A}
+B) {B}
+C) {C}
+D) {D}
+""".strip()
+
+MULTILINGUAL_ANSWER_REGEXES = [
+    r"Answer\s*:",
+    r"Answer\s*:​​​​​​",  # Korean invisible character
+    r"উত্তর\s*:",
+    r"उत्तर\s*:",
+    r"উত্তরঃ",
+    r"উত্তর\s*:",
+    r"Antwort\s*:",
+    r"답변\s*:",
+    r"정답\s*:",
+    r"답\s*:",
+    r"答案\s*：",
+    r"答案\s*:",
+    r"答\s*：",
+    r"答\s*:",
+    r"答复\s*：",
+    r"答曰\s*：",
+    r"الإجابة:",
+    r"الجواب:",
+    r"إجابة:",
+    r"الإجابة النهائية:",
+    r"الإجابة الصحيحة:",
+    r"الإجابة الصحيحة هي:",
+    r"الإجابة هي:",
+    r"Respuesta\s*:",
+    r"Risposta\s*:",
+    r"答え\s*:",
+    r"答え\s*：",
+    r"回答\s*:",
+    r"回答\s*：",
+    r"解答\s*:",
+    r"Jawaban\s*:",
+    r"Réponse\s*:",
+    r"Resposta\s*:",
+    r"Jibu\s*:",
+    r"Idahun\s*:",
+    r"Ìdáhùn\s*:",
+    r"Idáhùn\s*:",
+    r"Àmọ̀nà\s*:",
+    r"Àdáhùn\s*:",
+    r"Ànúgọ\s*:",
+    r"Àṣàyàn\s*:",
+]
+
+MULTILINGUAL_ANSWER_PATTERN_TEMPLATE = (
+    r"(?i){}\s*([A-D]|[أ-د]|[অ]|[ব]|[ড]|[ঢ]|[Ａ]|[Ｂ]|[Ｃ]|[Ｄ])"
+)
+
+
+def normalize_response(response: str) -> str:
+    """
+    Normalize the response by removing markdown and LaTeX formatting that may prevent a match.
+    """
+
+    return (
+        response.replace("**", "")
+        .replace("$\\boxed{", "")
+        .replace("}$", "")
+        .replace("\\$", "")
+        .replace("$\\text{", "")
+        .replace("$", "")
+        .replace("\\mathrm{", "")
+        .replace("\\{", "")
+        .replace("\\text", "")
+        .replace("\\(", "")
+        .replace("\\mathbf{", "")
+        .replace("{", "")
+        .replace("\\boxed", "")
+    )
+
+
+def normalize_extracted_answer(extracted_answer: str) -> str:
+    return (
+        # In arabic these are the letters used for A-D in multiple choice questions
+        extracted_answer.replace("أ", " A")
+        .replace("ب", " B")
+        .replace("ج", " C")
+        .replace("د", " D")
+        # In Bengali these are the letters used for A-D in multiple choice questions
+        .replace("অ", " A")
+        .replace("ব", " B")
+        .replace("ড", " C")
+        .replace("ঢ", " D")
+        # In Japanese these are the letters sometimes used for A-D in multiple choice questions
+        .replace("Ａ", " A")
+        .replace("Ｂ", " B")
+        .replace("Ｃ", " C")
+        .replace("Ｄ", " D")
+        .strip()
+    )
+
+
+class MMLUTask(BaseTask):
+    """
+    MMLU Task.
+    """
+
+    def __init__(self, dataset, *args, **kwargs):
+        super().__init__(dataset, *args, **kwargs)
+
+    def preprocess_sample(self, sample):
+        content = QUERY_TEMPLATE_MULTICHOICE.format(**sample)
+        return {
+            "role": "user",
+            "content": content,
+        }
+
+    def postprocess_sample(self, sample):
+        normalized = normalize_response(sample)
+        return normalized
+
+    def score_sample(self, sample, expected):
+        extracted_answer = None
+        for answer_regex in MULTILINGUAL_ANSWER_REGEXES:
+            regex = MULTILINGUAL_ANSWER_PATTERN_TEMPLATE.format(answer_regex)
+            match = re.search(regex, sample)
+            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
+        )
+        # TODO: generalize this into SingleEvalResult
+        return score
+
+    def aggregate_results(self, eval_results):
+        return EvaluateResponse(
+            metrics={"score": str(sum(eval_results) / len(eval_results))}
+        )