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

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,
 )
 
 
@@ -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="Retrieve and rank answers based on medical question " "understanding",
+                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",
+            ),
+        )

helm/benchmark/scenarios/mmlu_scenario.py

@@ -2,6 +2,7 @@ import csv
 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.common.hierarchical_logger import hlog
 from helm.benchmark.scenarios.scenario import (
@@ -14,6 +15,7 @@ from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     Input,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -124,3 +126,16 @@ class MMLUScenario(Scenario):
             instances.extend(self.process_csv(csv_path, splits[split]))
 
         return instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        return ScenarioMetadata(
+            name="mmlu",
+            display_name="MMLU (Massive Multitask Language Understanding)",
+            short_display_name="MMLU",
+            description="The Massive Multitask Language Understanding (MMLU) benchmark for "
+            "knowledge-intensive question answering across 57 domains [(Hendrycks et al., "
+            "2021)](https://openreview.net/forum?id=d7KBjmI3GmQ).",
+            taxonomy=TaxonomyInfo(task="question answering", what="?", when="?", who="?", language="English"),
+            main_metric="exact_match",
+            main_split="test",
+        )

helm/benchmark/scenarios/msmarco_scenario.py

@@ -4,6 +4,7 @@ import random
 from collections import defaultdict
 from typing import Dict, List, Tuple, Optional, Union
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.general import ensure_file_downloaded, ensure_directory_exists
 from helm.common.hierarchical_logger import hlog
 from helm.benchmark.scenarios.scenario import (
@@ -13,6 +14,7 @@ from helm.benchmark.scenarios.scenario import (
     TRAIN_SPLIT,
     VALID_SPLIT,
     CORRECT_TAG,
+    ScenarioMetadata,
     make_rank_tag,
     make_relevance_tag,
     Input,
@@ -657,3 +659,31 @@ class MSMARCOScenario(Scenario):
         valid_instances = self.get_valid_instances()
 
         return train_instances + valid_instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        if self.track == self.REGULAR_TRACK:
+            return ScenarioMetadata(
+                name="msmarco_regular",
+                display_name="MS MARCO (regular track)",
+                short_display_name="MS MARCO (regular)",
+                description="The MS MARCO benchmark's regular track for passage retrieval in information "
+                "retrieval "
+                "[(https://microsoft.github.io/msmarco/)](https://microsoft.github.io/msmarco/).",
+                taxonomy=TaxonomyInfo(task="information retrieval", what="?", when="?", who="?", language="English"),
+                main_metric="RR@10",
+                main_split="valid",
+            )
+        elif self.track == self.TREC_TRACK:
+            return ScenarioMetadata(
+                name="msmarco_trec",
+                display_name="MS MARCO (TREC track)",
+                short_display_name="MS MARCO (TREC)",
+                description="The MS MARCO benchmark's deep learning TREC track for passage retrieval in "
+                "information retrieval "
+                "[(https://trec.nist.gov)](https://microsoft.github.io/msmarco/).",
+                taxonomy=TaxonomyInfo(task="information retrieval", what="?", when="?", who="?", language="English"),
+                main_metric="NDCG@10",
+                main_split="valid",
+            )
+        else:
+            raise Exception(f"Unknown track {self.track}")

helm/benchmark/scenarios/mtsamples_procedures_scenario.py

@@ -1,6 +1,7 @@
 import os
 import requests
 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_directory_exists
 
@@ -142,3 +144,23 @@ class MTSamplesProceduresScenario(Scenario):
                 print(f"Error processing {file_name}: {e}")
 
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="mtsamples_procedures",
+            display_name="MTSamples Procedures",
+            description="MTSamples Procedures is a benchmark composed of transcribed operative notes, "
+            "focused on documenting surgical procedures. Each example presents a brief "
+            "patient case involving a surgical intervention, and the model is tasked with "
+            "generating a coherent and clinically accurate procedural summary or treatment "
+            "plan.",
+            taxonomy=TaxonomyInfo(
+                task="Text generation",
+                what="Document and extract information about medical procedures",
+                when="Post-procedure",
+                who="Clinician, Researcher",
+                language="English",
+            ),
+            main_metric="mtsamples_procedures_accuracy",
+            main_split="test",
+        )

helm/benchmark/scenarios/mtsamples_replicate_scenario.py

@@ -1,6 +1,7 @@
 import os
 import requests
 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_directory_exists
 
@@ -140,3 +142,23 @@ class MTSamplesReplicateScenario(Scenario):
                 print(f"Error processing {file_name}: {e}")
 
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="mtsamples_replicate",
+            display_name="MTSamples",
+            short_display_name="MTSamples",
+            description="MTSamples Replicate is a benchmark that provides transcribed medical reports "
+            "from various specialties. It is used to evaluate a model's ability to generate "
+            "clinically appropriate treatment plans based on unstructured patient "
+            "documentation [(MTSamples, 2025)](https://mtsamples.com).",
+            taxonomy=TaxonomyInfo(
+                task="Text generation",
+                what="Generate treatment plans based on clinical notes",
+                when="Post-diagnosis",
+                who="Clinician",
+                language="English",
+            ),
+            main_metric="mtsamples_replicate_accuracy",
+            main_split="test",
+        )

helm/benchmark/scenarios/n2c2_ct_matching_scenario.py

@@ -4,6 +4,7 @@ import re
 from typing import Any, Dict, List, Tuple, Optional
 import xml.etree.ElementTree as ET
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.general import ensure_directory_exists
 from helm.benchmark.scenarios.scenario import (
     TEST_SPLIT,
@@ -14,6 +15,7 @@ from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     Reference,
     Output,
+    ScenarioMetadata,
 )
 
 ORIGINAL_DEFINITIONS = {
@@ -275,3 +277,21 @@ class N2C2CTMatchingScenario(Scenario):
                     )
 
         return instances
+
+    def get_metadata(self):
+        return ScenarioMetadata(
+            name="n2c2_ct_matching",
+            display_name="N2C2-CT Matching",
+            short_display_name="N2C2-CT",
+            description="A dataset that provides clinical notes and asks the model to classify whether "
+            "the patient is a valid candidate for a provided clinical trial.",
+            taxonomy=TaxonomyInfo(
+                task="Classification",
+                what="Classify whether a patient is a valid candidate for a clinical " "trial based on clinical notes",
+                when="Pre-Trial",
+                who="Researcher",
+                language="English",
+            ),
+            main_metric="exact_match",
+            main_split="test",
+        )

helm/benchmark/scenarios/narrativeqa_scenario.py

@@ -3,6 +3,7 @@ import random
 import csv
 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,
@@ -13,6 +14,7 @@ from helm.benchmark.scenarios.scenario import (
     Input,
     PassageQuestionInput,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -155,3 +157,21 @@ class NarrativeQAScenario(Scenario):
             instances.extend(self.get_split_instances(summaries_file=summaries_file, qaps_file=qaps_file, split=split))
 
         return instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        return ScenarioMetadata(
+            name="narrative_qa",
+            display_name="NarrativeQA",
+            short_display_name=None,
+            description="The NarrativeQA benchmark for reading comprehension over narratives [(Kočiský "
+            "et al., 2017)](https://aclanthology.org/Q18-1023/).",
+            taxonomy=TaxonomyInfo(
+                task="question answering",
+                what="passages are books and movie scripts, questions are unknown",
+                when="?",
+                who="?",
+                language="English",
+            ),
+            main_metric="f1_score",
+            main_split="test",
+        )

helm/benchmark/scenarios/natural_qa_scenario.py

@@ -6,6 +6,7 @@ import os
 import re
 import html
 import random
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.hierarchical_logger import htrack_block, hlog
 from typing import List, Dict
 
@@ -20,6 +21,7 @@ from helm.benchmark.scenarios.scenario import (
     PassageQuestionInput,
     Input,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -324,3 +326,33 @@ class NaturalQAScenario(Scenario):
             instances.extend(self.get_file_instances(target_path))
 
         return instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        if self.context_mode == "closedbook":
+            name = "natural_qa_closedbook"
+            display_name = "NaturalQuestions (closed-book)"
+        elif self.context_mode == "openbook_longans":
+            name = "natural_qa_openbook_longans"
+            display_name = "NaturalQuestions (open-book)"
+        elif self.context_mode == "openbook_wiki":
+            name = "natural_qa_openbook_wiki"
+            display_name = "NaturalQuestions (open-book Wiki)"
+        else:
+            raise Exception(f"Unknown context_mode {self.context_mode}")
+        return ScenarioMetadata(
+            name=name,
+            display_name=display_name,
+            description="The NaturalQuestions [(Kwiatkowski et al., "
+            "2019)](https://aclanthology.org/Q19-1026/) benchmark for question answering "
+            "based on naturally-occurring queries through Google Search. The input does not "
+            "include the Wikipedia page with the answer.",
+            taxonomy=TaxonomyInfo(
+                task="question answering",
+                what="passages from Wikipedia, questions from search queries",
+                when="2010s",
+                who="web users",
+                language="English",
+            ),
+            main_metric="f1_score",
+            main_split="valid",
+        )