crfm-helm 0.5.7__py3-none-any.whl → 0.5.8__py3-none-any.whl

helm/benchmark/scenarios/legal_support_scenario.py

@@ -2,6 +2,7 @@ import json
 import os
 from typing import List
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.general import ensure_file_downloaded
 from helm.benchmark.scenarios.scenario import (
     Scenario,
@@ -13,6 +14,7 @@ from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     Input,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -102,3 +104,14 @@ class LegalSupportScenario(Scenario):
                 instances.append(instance)
 
         return instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        return ScenarioMetadata(
+            name="legal_support",
+            display_name="LegalSupport",
+            description="Scenario introduced in this work to measure fine-grained legal reasoning "
+            "through reverse entailment.",
+            taxonomy=TaxonomyInfo(task="?", what="n/a", when="n/a", who="n/a", language="synthetic"),
+            main_metric="quasi_exact_match",
+            main_split="test",
+        )

helm/benchmark/scenarios/legalbench_scenario.py

@@ -5,6 +5,7 @@ import datasets
 from pathlib import Path
 from typing import List, Dict
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.general import ensure_file_downloaded, ensure_directory_exists
 from helm.benchmark.scenarios.scenario import (
     Scenario,
@@ -15,6 +16,7 @@ from helm.benchmark.scenarios.scenario import (
     TEST_SPLIT,
     Input,
     Output,
+    ScenarioMetadata,
 )
 
 PROMPT_SETTINGS_URL = "https://raw.githubusercontent.com/HazyResearch/legalbench/main/helm_prompt_settings.jsonl"
@@ -144,3 +146,21 @@ class LegalBenchScenario(Scenario):
                 instances.append(instance)
 
         return instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        return ScenarioMetadata(
+            name="legalbench",
+            display_name="LegalBench",
+            description="LegalBench is a large collaboratively constructed benchmark of legal "
+            "reasoning. Five representative tasks are included here. See [(Guha et al, "
+            "2023)[https://arxiv.org/abs/2308.11462] for more details.",
+            taxonomy=TaxonomyInfo(
+                task="text classification",
+                what="fact patterns, questions, and legal documents",
+                when="n/a",
+                who="lawyers",
+                language="English",
+            ),
+            main_metric="quasi_exact_match",
+            main_split="test",
+        )

helm/benchmark/scenarios/lex_glue_scenario.py

@@ -16,6 +16,7 @@ from helm.benchmark.scenarios.scenario import (
     TEST_SPLIT,
     Input,
     Output,
+    ScenarioMetadata,
 )
 
 ECTHR_A = "ecthr_a"
@@ -261,3 +262,13 @@ class LexGLUEScenario(Scenario):
         for subset in self.subsets:
             instances.extend(self.get_instances_for_subset(subset, output_path))
         return instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        return ScenarioMetadata(
+            name="lex_glue",
+            display_name="LexGLUE",
+            description="A Benchmark Dataset for Legal Language Understanding in English",
+            taxonomy=None,
+            main_metric="classification_macro_f1",
+            main_split="test",
+        )

helm/benchmark/scenarios/lextreme_scenario.py

@@ -16,6 +16,7 @@ from helm.benchmark.scenarios.scenario import (
     TEST_SPLIT,
     Output,
     Input,
+    ScenarioMetadata,
 )
 
 
@@ -466,3 +467,13 @@ class LEXTREMEScenario(Scenario):
         for subset in self.subsets:
             instances.extend(self.get_instances_for_subset(subset, output_path))
         return instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        return ScenarioMetadata(
+            name="lextreme",
+            display_name="LEXTREME",
+            description="A Multilingual Legal Benchmark for Natural Language Understanding",
+            taxonomy=None,
+            main_metric="classification_macro_f1",
+            main_split="test",
+        )

helm/benchmark/scenarios/lsat_qa_scenario.py

@@ -2,6 +2,7 @@ import os
 import json
 from typing import Dict, List
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.general import ensure_file_downloaded, ensure_directory_exists
 from helm.benchmark.scenarios.scenario import (
     Scenario,
@@ -13,6 +14,7 @@ from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     PassageQuestionInput,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -143,3 +145,15 @@ class LSATScenario(Scenario):
                             instances.append(instance)
 
         return instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        return ScenarioMetadata(
+            name="lsat_qa",
+            display_name="LSAT",
+            description="The LSAT benchmark for measuring analytical reasoning on the Law School "
+            "Admission Test (LSAT; [Zhong et al., "
+            "2021](https://arxiv.org/pdf/2104.06598.pdf)).",
+            taxonomy=TaxonomyInfo(task="?", what="n/a", when="n/a", who="n/a", language="synthetic"),
+            main_metric="quasi_exact_match",
+            main_split="test",
+        )

helm/benchmark/scenarios/madinah_qa_scenario.py

@@ -0,0 +1,73 @@
+import os
+from typing import Dict, List
+
+import datasets
+
+from helm.common.general import ensure_directory_exists
+from helm.benchmark.scenarios.scenario import (
+    Scenario,
+    Instance,
+    Reference,
+    TEST_SPLIT,
+    TRAIN_SPLIT,
+    CORRECT_TAG,
+    Input,
+    Output,
+)
+
+
+class MadinahQAScenario(Scenario):
+    """MadinahQA Scenario"""
+
+    name = "madinah_qa"
+    description = "Arabic language competency benchmark"
+    tags = ["language", "multiple_choice"]
+
+    OPTIONS = ["A", "B", "C", "D"]
+    HF_SPLIT_TO_HELM_SPLIT = {"dev": TRAIN_SPLIT, "test": TEST_SPLIT}
+    SUBSETS = ["Arabic Language (General)", "Arabic Language (Grammar)"]
+
+    def __init__(self, subset: str):
+        super().__init__()
+        subset = subset.replace("_", " ")
+        if subset not in self.SUBSETS:
+            raise Exception(f"Unknown subset: {subset}")
+        self.subset = subset
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        cache_dir = os.path.join(output_path, "data")
+        ensure_directory_exists(cache_dir)
+        instances: List[Instance] = []
+        dataset_splits: Dict[str, datasets.Dataset] = datasets.load_dataset(
+            "MBZUAI/MadinahQA",
+            self.subset,
+            revision="62e7c86ac5c07245a5a952722691d77ddb41f695",
+            cache_dir=cache_dir,
+        )
+
+        # Read all instances
+        for split_name, dataset in dataset_splits.items():
+            assert isinstance(dataset, datasets.Dataset)
+            for row_index, row in enumerate(dataset):
+                input = Input(text=row["Question"])
+                references: List[Reference] = []
+                correct_option_index = ord(row["Answer Key"]) - ord("A") + 1
+                for option_index in range(1, 6):
+                    column_name = f"Option {option_index}"
+                    if not row[column_name]:
+                        continue
+                    references.append(
+                        Reference(
+                            output=Output(text=row[column_name]),
+                            tags=[CORRECT_TAG] if option_index == correct_option_index else [],
+                        )
+                    )
+                instance = Instance(
+                    id=f"id{row_index}",
+                    input=input,
+                    references=references,
+                    split=self.HF_SPLIT_TO_HELM_SPLIT[split_name],
+                )
+                instances.append(instance)
+
+        return instances

helm/benchmark/scenarios/math_scenario.py

@@ -4,6 +4,7 @@ import typing
 from typing import Dict, List, Optional
 from datasets import load_dataset, DatasetDict
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.general import ensure_directory_exists
 from helm.benchmark.scenarios.scenario import (
     Scenario,
@@ -14,6 +15,7 @@ from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     Input,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -450,3 +452,27 @@ class MATHScenario(Scenario):
                 instances.append(instance)
 
         return instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        if self.use_chain_of_thought:
+            return ScenarioMetadata(
+                name="math_chain_of_thought",
+                display_name="MATH (chain-of-thought)",
+                description="The MATH benchmark for measuring mathematical problem solving on competition "
+                "math problems with chain-of-thought style reasoning [(Hendrycks et al., "
+                "2021)](https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/hash/be83ab3ecd0db773eb2dc1b0a17836a1-Abstract-round2.html).",
+                taxonomy=TaxonomyInfo(task="?", what="n/a", when="n/a", who="n/a", language="synthetic"),
+                main_metric="math_equiv_chain_of_thought",
+                main_split="test",
+            )
+        else:
+            return ScenarioMetadata(
+                name="math_regular",
+                display_name="MATH",
+                description="The MATH benchmark for measuring mathematical problem solving on competition "
+                "math problems [(Hendrycks et al., "
+                "2021)](https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/hash/be83ab3ecd0db773eb2dc1b0a17836a1-Abstract-round2.html).",
+                taxonomy=TaxonomyInfo(task="?", what="n/a", when="n/a", who="n/a", language="synthetic"),
+                main_metric="math_equiv",
+                main_split="test",
+            )

helm/benchmark/scenarios/mbzuai_human_translated_arabic_mmlu.py

@@ -0,0 +1,68 @@
+import os
+from typing import List
+
+import datasets
+
+from helm.common.general import ensure_directory_exists
+from helm.benchmark.scenarios.scenario import (
+    Scenario,
+    Instance,
+    Reference,
+    TEST_SPLIT,
+    CORRECT_TAG,
+    Input,
+    Output,
+)
+
+
+class MBZUAIHumanTranslatedArabicMMLUScenario(Scenario):
+    """MBZUAI Human-Translated Arabic MMLU
+
+    A translation from MBZUAI by human translators of the Massive Multitask Language Understanding benchmark from this paper:
+
+    - https://arxiv.org/pdf/2009.03300.pdf
+    """  # noqa: E501
+
+    name = "mbzuai_human_translated_arabic_mmlu"
+    description = (
+        "A translation from MBZUAI by human translators of the Massive Multitask Language Understanding benchmark"
+    )
+    tags = ["knowledge", "multiple_choice"]
+
+    def __init__(self, subject: str):
+        super().__init__()
+        self.subject: str = subject
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        cache_dir = os.path.join(output_path, "data")
+        ensure_directory_exists(cache_dir)
+        dataset = datasets.load_dataset(
+            "MBZUAI/human_translated_arabic_mmlu",
+            self.subject,
+            revision="5ed7830fd678cfa6f2d7f0a1a13a4e1a1fa422ac",
+            cache_dir=cache_dir,
+            split="test",
+        )
+        assert isinstance(dataset, datasets.Dataset)
+
+        # Read all instances
+        instances: List[Instance] = []
+        for row_index, row in enumerate(dataset):
+            input = Input(text=row["question"])
+            references: List[Reference] = []
+            for choice_index, choice in enumerate(row["choices"]):
+                references.append(
+                    Reference(
+                        output=Output(text=choice),
+                        tags=[CORRECT_TAG] if choice_index == row["answer"] else [],
+                    )
+                )
+            instance = Instance(
+                id=f"id-{self.subject}-{row_index}",
+                input=input,
+                references=references,
+                split=TEST_SPLIT,
+            )
+            instances.append(instance)
+
+        return instances

helm/benchmark/scenarios/med_dialog_scenario.py

@@ -2,8 +2,18 @@ import json
 import os
 from typing import List
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.general import ensure_directory_exists, ensure_file_downloaded
-from helm.benchmark.scenarios.scenario import Scenario, Instance, Reference, ALL_SPLITS, CORRECT_TAG, Input, Output
+from helm.benchmark.scenarios.scenario import (
+    Scenario,
+    Instance,
+    Reference,
+    ALL_SPLITS,
+    CORRECT_TAG,
+    Input,
+    Output,
+    ScenarioMetadata,
+)
 
 
 class MedDialogScenario(Scenario):
@@ -133,3 +143,24 @@ class MedDialogScenario(Scenario):
                         )
 
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="med_dialog",
+            display_name="MedDialog",
+            short_display_name="MedDialog",
+            description="MedDialog is a benchmark of real-world doctor-patient conversations focused on "
+            "health-related concerns and advice. Each dialogue is paired with a "
+            "one-sentence summary that reflects the core patient question or exchange. The "
+            "benchmark evaluates a model's ability to condense medical dialogue into "
+            "concise, informative summaries.",
+            taxonomy=TaxonomyInfo(
+                task="Text generation",
+                what="Generate summaries of doctor-patient conversations",
+                when="Any",
+                who="Clinician",
+                language="English",
+            ),
+            main_metric="med_dialog_accuracy",
+            main_split="test",
+        )

helm/benchmark/scenarios/med_mcqa_scenario.py

@@ -2,6 +2,7 @@ import json
 import os
 from typing import Dict, List
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.general import ensure_file_downloaded
 from helm.benchmark.scenarios.scenario import (
     Scenario,
@@ -12,6 +13,7 @@ from helm.benchmark.scenarios.scenario import (
     VALID_SPLIT,
     Input,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -109,3 +111,15 @@ class MedMCQAScenario(Scenario):
                     instances.append(instance)
 
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="med_mcqa",
+            display_name="MedMCQA",
+            description='MedMCQA is a "multiple-choice question answering (MCQA) dataset designed to '
+            "address real-world medical entrance exam questions ([Flores et al. "
+            "2020](https://arxiv.org/abs/2203.14371)).",
+            taxonomy=TaxonomyInfo(task="question answering", what="n/a", when="n/a", who="n/a", language="English"),
+            main_metric="exact_match",
+            main_split="valid",
+        )

helm/benchmark/scenarios/med_qa_scenario.py

@@ -2,6 +2,7 @@ import json
 import os
 from typing import Dict, List
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.general import ensure_file_downloaded
 from helm.benchmark.scenarios.scenario import (
     Scenario,
@@ -12,6 +13,7 @@ from helm.benchmark.scenarios.scenario import (
     VALID_SPLIT,
     Input,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -103,3 +105,15 @@ class MedQAScenario(Scenario):
                     instances.append(instance)
 
         return instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        return ScenarioMetadata(
+            name="med_qa",
+            display_name="MedQA",
+            description="MedQA is an open domain question answering dataset composed of questions from "
+            "professional medical board exams ([Jin et al. "
+            "2020](https://arxiv.org/pdf/2009.13081.pdf)).",
+            taxonomy=TaxonomyInfo(task="question answering", what="n/a", when="n/a", who="n/a", language="English"),
+            main_metric="quasi_exact_match",
+            main_split="test",
+        )

helm/benchmark/scenarios/medalign_scenario.py

@@ -1,5 +1,6 @@
 from typing import List
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.benchmark.scenarios.scenario import (
     Scenario,
     Instance,
@@ -8,6 +9,7 @@ from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     PassageQuestionInput,
     Output,
+    ScenarioMetadata,
 )
 from helm.benchmark.scenarios.medalign_scenario_helper import return_dataset_dataframe  # type: ignore
 
@@ -92,3 +94,24 @@ class MedalignScenario(Scenario):
     def get_instances(self, output_path: str) -> List[Instance]:
         dataset = return_dataset_dataframe(self.max_length, self.data_path)
         return self.process_tsv(dataset)
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="medalign",
+            display_name="MedAlign",
+            short_display_name="MedAlign",
+            description="MedAlign is a benchmark that evaluates a model's ability to interpret and "
+            "follow instructions grounded in longitudinal electronic health records (EHR). "
+            "Each instance includes an event-stream style patient record and a natural "
+            "language question or task, requiring clinically informed reading comprehension "
+            "and reasoning [(Fleming et al., 2023)](https://arxiv.org/abs/2308.14089).",
+            taxonomy=TaxonomyInfo(
+                task="Text generation",
+                what="Answer questions and follow instructions over longitudinal EHR",
+                when="Any",
+                who="Clinician, Researcher",
+                language="English",
+            ),
+            main_metric="medalign_accuracy",
+            main_split="test",
+        )

helm/benchmark/scenarios/medbullets_scenario.py

@@ -3,6 +3,7 @@ import csv
 import sys
 from typing import List
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     TEST_SPLIT,
@@ -11,6 +12,7 @@ from helm.benchmark.scenarios.scenario import (
     Output,
     Reference,
     Scenario,
+    ScenarioMetadata,
 )
 from helm.common.general import ensure_file_downloaded
 
@@ -143,3 +145,23 @@ class MedBulletsScenario(Scenario):
             csv_path = self.download_csv(output_path, split_suffix)
             instances.extend(self.process_csv(csv_path, split))
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="medbullets",
+            display_name="Medbullets",
+            description="Medbullets is a benchmark of USMLE-style medical questions designed to assess "
+            "a model's ability to understand and apply clinical knowledge. Each question is "
+            "accompanied by a patient scenario and five multiple-choice options, similar to "
+            "those found on Step 2 and Step 3 board exams [(MedBullets, "
+            "2025)](https://step2.medbullets.com).",
+            taxonomy=TaxonomyInfo(
+                task="Question answering",
+                what="Medical knowledge testing",
+                when="Any",
+                who="Medical student, . Researcher",
+                language="English",
+            ),
+            main_metric="exact_match",
+            main_split="test",
+        )

helm/benchmark/scenarios/medcalc_bench_scenario.py

@@ -1,6 +1,7 @@
 from typing import Dict, List
 from datasets import load_dataset
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.hierarchical_logger import hlog
 from helm.benchmark.scenarios.scenario import (
     Scenario,
@@ -10,6 +11,7 @@ from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     PassageQuestionInput,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -125,3 +127,23 @@ class MedCalcBenchScenario(Scenario):
             instances.extend(self.process_csv(data, split))
 
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="medcalc_bench",
+            display_name="MedCalc-Bench",
+            description="MedCalc-Bench is a benchmark designed to evaluate models on their ability to "
+            "compute clinically relevant values from patient notes. Each instance consists "
+            "of a clinical note describing the patient's condition, a diagnostic question "
+            "targeting a specific medical value, and a ground truth response. [(Khandekar "
+            "et al., 2024)](https://arxiv.org/abs/2406.12036).",
+            taxonomy=TaxonomyInfo(
+                task="Computational reasoning",
+                what="Compute a specific medical value from a patient note",
+                when="Any",
+                who="Clinician, Researcher",
+                language="English",
+            ),
+            main_metric="medcalc_bench_accuracy",
+            main_split="test",
+        )

helm/benchmark/scenarios/medec_scenario.py

@@ -1,6 +1,7 @@
 import csv
 import os
 from typing import List
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.benchmark.scenarios.scenario import (
     Scenario,
     Instance,
@@ -9,6 +10,7 @@ from helm.benchmark.scenarios.scenario import (
     TEST_SPLIT,
     Input,
     Output,
+    ScenarioMetadata,
 )
 from helm.common.general import ensure_file_downloaded
 
@@ -123,3 +125,24 @@ class MedecScenario(Scenario):
         instances.extend(self.process_csv(test_csv, TEST_SPLIT))
 
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="medec",
+            display_name="Medec",
+            description="Medec is a benchmark composed of clinical narratives that include either "
+            "correct documentation or medical errors. Each entry includes sentence-level "
+            "identifiers and an associated correction task. The model must review the "
+            "narrative and either identify the erroneous sentence and correct it, or "
+            "confirm that the text is entirely accurate [(Abacha et al., "
+            "2025)](https://arxiv.org/abs/2412.19260).",
+            taxonomy=TaxonomyInfo(
+                task="Classification",
+                what="Detect and correct errors in medical narratives",
+                when="Any",
+                who="Researcher, Clinician",
+                language="English",
+            ),
+            main_metric="medec_error_flag_accuracy",
+            main_split="test",
+        )

helm/benchmark/scenarios/medhallu_scenario.py

@@ -1,6 +1,7 @@
 from typing import List
 from datasets import load_dataset
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.benchmark.scenarios.scenario import (
     Scenario,
     Instance,
@@ -9,6 +10,7 @@ from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     Output,
     Input,
+    ScenarioMetadata,
 )
 
 
@@ -70,3 +72,24 @@ Answer: {answer}
             )
             instances.append(hallucinated_instance)
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="medhallu",
+            display_name="MedHallu",
+            description="MedHallu is a benchmark focused on evaluating factual correctness in "
+            "biomedical question answering. Each instance contains a PubMed-derived "
+            "knowledge snippet, a biomedical question, and a model-generated answer. The "
+            "task is to classify whether the answer is factually correct or contains "
+            "hallucinated (non-grounded) information. This benchmark is designed to assess "
+            "the factual reliability of medical language models.",
+            taxonomy=TaxonomyInfo(
+                task="Classification",
+                what="Verify whether answers to questions from PubMed articles are " "factual or hallucinated",
+                when="Any",
+                who="Researcher",
+                language="English",
+            ),
+            main_metric="exact_match",
+            main_split="test",
+        )

helm/benchmark/scenarios/medhelm/judges.yaml

@@ -0,0 +1,14 @@
+# The judges to be used for evaluating the note summary scenario.
+# name: The short name for the judge.
+# model: The field value matching the 'model_name' field under model_deployments.yaml
+# model_deployment: The field value matching the 'name' under model_deployments.yaml.
+judges:
+  - name: "gpt"
+    model: "openai/gpt-4o-2024-05-13"
+    model_deployment: "stanfordhealthcare/gpt-4o-2024-05-13"
+  - name: "llama"
+    model: "meta/llama-3.3-70b-instruct"
+    model_deployment: "stanfordhealthcare/llama-3.3-70b-instruct"
+  - name: "claude"
+    model: "anthropic/claude-3-7-sonnet-20250219"
+    model_deployment: "stanfordhealthcare/claude-3-7-sonnet-20250219"