crfm-helm 0.5.7__py3-none-any.whl → 0.5.9__py3-none-any.whl

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 +105,21 @@ 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="multiple-choice question answering",
+                what="US medical licensing exams",
+                when="before 2020",
+                who="problem setters",
+                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"

helm/benchmark/scenarios/medhelm_configurable_scenario.py

@@ -0,0 +1,101 @@
+import string
+import json
+import pandas as pd
+from typing import List
+
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
+from helm.benchmark.run_specs.medhelm.benchmark_config import get_benchmark_config_from_path
+from helm.common.general import check_file_exists
+
+from helm.benchmark.scenarios.scenario import (
+    Scenario,
+    Instance,
+    CORRECT_TAG,
+    Reference,
+    Input,
+    Output,
+    TEST_SPLIT,
+    ScenarioMetadata,
+)
+
+
+class MedHELMConfigurableScenario(Scenario):
+    """
+    MedHELM configuratble scenario
+    """
+
+    tags = ["biomedical"]
+
+    def __init__(self, name: str, config_path: str):
+        super().__init__()
+        self.benchmark_config = get_benchmark_config_from_path(config_path)
+        self.name = name
+        self.description = self.benchmark_config.description
+
+    def get_columns_in_template(self, template: str) -> List[str]:
+        """
+        Extract field names from a template string using Python's Formatter.
+        Example: "Name: {name}, Age: {age}" → ["name", "age"]
+        """
+        formatter = string.Formatter()
+        return [fname for _, fname, _, _ in formatter.parse(template) if fname]
+
+    def populate_template(self, template: str, row: pd.Series, fields: List[str]) -> str:
+        """
+        Populate the template with values from the row using format_map.
+        Missing fields default to empty string.
+        """
+        mapping = {field: row.get(field, "") for field in fields}
+        return template.format_map(mapping)
+
+    def get_references(self, row: pd.Series) -> List[Reference]:
+        references: List[Reference] = []
+        if "correct_answer" in row:
+            references.append(Reference(Output(text=row["correct_answer"]), tags=[CORRECT_TAG]))
+        if "incorrect_answers" in row:
+            for incorrect_answer in row["incorrect_answers"]:
+                references.append(Reference(Output(text=incorrect_answer), tags=[]))
+        return references
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        check_file_exists(self.benchmark_config.prompt_file, msg=f"Prompt file for {self.name} does not exist")
+        check_file_exists(self.benchmark_config.dataset_file, msg=f"Dataset file for {self.name} does not exist")
+        instances: List[Instance] = []
+        df = pd.read_csv(self.benchmark_config.dataset_file)
+        if "correct_answer" not in df.columns:
+            if not self._is_llm_as_judge() or len(self.benchmark_config.metrics) > 1:
+                raise ValueError(
+                    "Dataset must contain 'correct_answer' column unless using jury_score as the only metric."
+                )
+        if "incorrect_answers" in df.columns:
+            df["incorrect_answers"] = df["incorrect_answers"].apply(json.loads)
+        with open(self.benchmark_config.prompt_file, "r") as f:
+            template = f.read()
+        fields = self.get_columns_in_template(template)
+        for _, row in df.iterrows():
+            filled = self.populate_template(template, row, fields)
+            prompt = Input(text=filled)
+            instances.append(Instance(input=prompt, references=self.get_references(row), split=TEST_SPLIT))
+        return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name=self.name,
+            display_name=self.name,
+            description=self.description,
+            taxonomy=TaxonomyInfo(
+                task="",
+                what="",
+                when="",
+                who="",
+                language="",
+            ),
+            main_metric=self.benchmark_config.main_metric.name,
+            main_split="test",
+        )
+
+    def _is_llm_as_judge(self) -> bool:
+        for metric in self.benchmark_config.metrics:
+            if metric.name == "jury_score":
+                return True
+        return False

helm/benchmark/scenarios/medi_qa_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,
     Input,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -49,7 +51,7 @@ class MediQAScenario(Scenario):
 
     name = "medi_qa"
     description = (
-        "MEDIQA is a benchmark designed to evaluate a model's ability to retrieve and generate"
+        "MEDIQA is a benchmark designed to evaluate a model's ability to generate"
         "medically accurate answers to patient-generated questions. Each instance includes a"
         "consumer health question, a set of candidate answers (used in ranking tasks), relevance"
         "annotations, and optionally, additional context. The benchmark focuses on supporting"
@@ -109,3 +111,24 @@ class MediQAScenario(Scenario):
             instances.extend(self.process_csv(data, split))
 
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="medi_qa",
+            display_name="MEDIQA",
+            description="MEDIQA is a benchmark designed to evaluate a model's ability to retrieve and "
+            "generate medically accurate answers to patient-generated questions. Each "
+            "instance includes a consumer health question, a set of candidate answers (used "
+            "in ranking tasks), relevance annotations, and optionally, additional context. "
+            "The benchmark focuses on supporting patient understanding and accessibility in "
+            "health communication.",
+            taxonomy=TaxonomyInfo(
+                task="Text generation",
+                what="Generate medically accurate answers to patient-generated questions.",
+                when="Any",
+                who="Clinician, Medical Student",
+                language="English",
+            ),
+            main_metric="medi_qa_accuracy",
+            main_split="test",
+        )

helm/benchmark/scenarios/medication_qa_scenario.py

@@ -3,9 +3,19 @@ from typing import List
 
 import pandas as pd
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.general import ensure_file_downloaded
 
-from helm.benchmark.scenarios.scenario import CORRECT_TAG, TEST_SPLIT, Input, Instance, Output, Reference, Scenario
+from helm.benchmark.scenarios.scenario import (
+    CORRECT_TAG,
+    TEST_SPLIT,
+    Input,
+    Instance,
+    Output,
+    Reference,
+    Scenario,
+    ScenarioMetadata,
+)
 
 
 class MedicationQAScenario(Scenario):
@@ -64,3 +74,23 @@ class MedicationQAScenario(Scenario):
         ]
 
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="medication_qa",
+            display_name="MedicationQA",
+            description="MedicationQA is a benchmark composed of open-ended consumer health questions "
+            "specifically focused on medications. Each example consists of a free-form "
+            "question and a corresponding medically grounded answer. The benchmark "
+            "evaluates a model's ability to provide accurate, accessible, and informative "
+            "medication-related responses for a lay audience.",
+            taxonomy=TaxonomyInfo(
+                task="Question answering",
+                what="Answer consumer medication-related questions",
+                when="Any",
+                who="Patient, Pharmacist",
+                language="English",
+            ),
+            main_metric="medication_qa_accuracy",
+            main_split="test",
+        )

helm/benchmark/scenarios/mental_health_scenario.py

@@ -1,5 +1,6 @@
 import pandas as pd
 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.common.general import check_file_exists
 
@@ -121,3 +123,24 @@ class MentalHealthScenario(Scenario):
         instances = self.process_dialogue_data(dialogue_data)
 
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="mental_health",
+            display_name="MentalHealth",
+            description="MentalHealth is a benchmark focused on evaluating empathetic communication in "
+            "mental health counseling. It includes real or simulated conversations between "
+            "patients and counselors, where the task is to generate compassionate and "
+            "appropriate counselor responses. The benchmark assesses a model's ability to "
+            "support patients emotionally and meaningfully engage in therapeutic "
+            "conversations.",
+            taxonomy=TaxonomyInfo(
+                task="Text generation",
+                what="Generate empathetic counseling responses in mental health " "conversations",
+                when="Any",
+                who="Counselors, Patients",
+                language="English",
+            ),
+            main_metric="mental_health_accuracy",
+            main_split="test",
+        )

helm/benchmark/scenarios/mimic_bhc_scenario.py

@@ -1,6 +1,7 @@
 import json
 from typing import Dict, List
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.general import check_file_exists
 from helm.benchmark.scenarios.scenario import (
     Input,
@@ -10,6 +11,7 @@ from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     Reference,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -101,3 +103,25 @@ class MIMICBHCScenario(Scenario):
                 )
 
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="mimic_bhc",
+            display_name="MIMIC-IV-BHC",
+            short_display_name="MIMIC-BHC",
+            description="MIMIC-BHC is a benchmark focused on summarization of discharge notes into "
+            "Brief Hospital Course (BHC) sections. It consists of curated discharge notes "
+            "from MIMIC-IV, each paired with its corresponding BHC summary. The benchmark "
+            "evaluates a model's ability to condense detailed clinical information into "
+            "accurate, concise summaries that reflect the patient's hospital stay [(Aali et "
+            "al., 2024)](https://doi.org/10.1093/jamia/ocae312).",
+            taxonomy=TaxonomyInfo(
+                task="Text generation",
+                what="Summarize the clinical note into a brief hospital course",
+                when="Upon hospital discharge",
+                who="Clinician",
+                language="English",
+            ),
+            main_metric="mimic_bhc_accuracy",
+            main_split="test",
+        )

helm/benchmark/scenarios/mimic_rrs_scenario.py

@@ -1,6 +1,7 @@
 import os
 from typing import Dict, List
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.general import check_file_exists
 from helm.benchmark.scenarios.scenario import (
     Input,
@@ -10,6 +11,7 @@ from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     Reference,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -96,3 +98,24 @@ class MIMICRRSScenario(Scenario):
             lines = file.readlines()
         lines = [line.strip() for line in lines]
         return lines
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="mimic_rrs",
+            display_name="MIMIC-RRS",
+            short_display_name="MIMIC-RRS",
+            description="MIMIC-RRS is a benchmark constructed from radiology reports in the MIMIC-III "
+            "database. It contains pairs of ‘Findings‘ and ‘Impression‘ sections, enabling "
+            "evaluation of a model's ability to summarize diagnostic imaging observations "
+            "into concise, clinically relevant conclusions [(Chen et al., "
+            "2023)](https://arxiv.org/abs/2211.08584).",
+            taxonomy=TaxonomyInfo(
+                task="Text generation",
+                what="Generate radiology report summaries from findings sections",
+                when="Post-imaging",
+                who="Radiologist",
+                language="English",
+            ),
+            main_metric="mimic_rrs_accuracy",
+            main_split="test",
+        )

helm/benchmark/scenarios/mimiciv_billing_code_scenario.py

@@ -2,6 +2,7 @@ import pandas as pd
 import numpy as np
 from typing import List
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.general import check_file_exists
 from helm.benchmark.scenarios.scenario import (
     Input,
@@ -11,6 +12,7 @@ from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     Reference,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -75,3 +77,23 @@ class MIMICIVBillingCodeScenario(Scenario):
                 continue
 
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="mimiciv_billing_code",
+            display_name="MIMIC-IV Billing Code",
+            description="MIMIC-IV Billing Code is a benchmark derived from discharge summaries in the "
+            "MIMIC-IV database, paired with their corresponding ICD-10 billing codes. The "
+            "task requires models to extract structured billing codes based on free-text "
+            "clinical notes, reflecting real-world hospital coding tasks for financial "
+            "reimbursement.",
+            taxonomy=TaxonomyInfo(
+                task="Classification",
+                what="Predict ICD-10 billing codes from clinical discharge notes",
+                when="During or after patient discharge",
+                who="Hospital Admistrator",
+                language="English",
+            ),
+            main_metric="mimiciv_billing_code_f1",
+            main_split="test",
+        )

helm/benchmark/scenarios/mmlu_pro_scenario.py

@@ -2,6 +2,7 @@ from typing import Dict, List
 from datasets import Dataset, load_dataset
 
 from helm.common.hierarchical_logger import hlog
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.benchmark.scenarios.scenario import (
     Scenario,
     Instance,
@@ -11,6 +12,7 @@ from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     Input,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -93,3 +95,19 @@ class MMLUProScenario(Scenario):
             instances.extend(self.process_dataset(data, split))
 
         return instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        return ScenarioMetadata(
+            name=self.name,
+            display_name="MMLU-Pro",
+            description=self.description,
+            main_metric="chain_of_thought_correctness",
+            main_split="test",
+            taxonomy=TaxonomyInfo(
+                task="question answering",
+                what="graduate-level questions in biology, physics, and chemistry",
+                who="domain experts",
+                when="2023",
+                language="English",
+            ),
+        )