crfm-helm 0.5.6__py3-none-any.whl → 0.5.10__py3-none-any.whl

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 +105,21 @@ 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/medalign_scenario_helper.py

@@ -2,22 +2,13 @@
 # type: ignore
 # fmt: off
 
-import ast
-import datetime
 import transformers
-import langchain
-import langchain.prompts
-import lxml.etree
 import os
 import pandas as pd
-import re
 import tiktoken
 
-from langchain_community.retrievers import BM25Retriever
 from tqdm import tqdm
-from typing import Any, Dict, Optional, Union, Callable
-from langchain.schema import Document
-import langchain_community
+from typing import Any, Dict, Optional, Callable
 
 from helm.common.general import check_file_exists
 
@@ -167,102 +158,13 @@ def get_tokenizer(tokenizer_name: str) -> Callable:
     return transformers.AutoTokenizer.from_pretrained(tokenizer_name, legacy=False)
 
 
-def retrieve_most_relevant_visits(ehr_visit_strs, query, target_length, tokenizer):
-    """
-    Retrieve and filter relevant EHR visits based on a query and target length.
-
-    This function retrieves electronic health record (EHR) visit strings, sorts them
-    by relevance using the BM25Retriever, and constructs a list of final documents
-    that fit within a specified character length. The final list ensures that the
-    most important visit isn't cut off and is sorted chronologically.
-
-    Parameters:
-        ehr_visit_strs (list of str): List of EHR visit strings.
-        query (str): Query string to retrieve relevant visits.
-        target_length (int): Maximum total token count for the final list of documents.
-        tokenizer (Callable): Tokenizer that converts text to tokens (used for tracking context length)
-
-    Returns:
-        list[str]: List of EHR visit strings sorted chronologically and constrained by the target length.
-    """
-    ehr_visits=re.split(r'(?=</encounter>\n)',ehr_visit_strs)
-    langchain_docs = [
-        langchain.schema.Document(page_content=doc) for doc in ehr_visits #broken since ehr_visit_strs is one string of all visits
-    ]
-    # `k` is the number of documents to retrieve
-    # We retrieve everything and just use the BM25Retriever to sort the documents
-    retriever = langchain_community.retrievers.BM25Retriever.from_documents(
-        langchain_docs, k=len(langchain_docs)
-    )
-
-    # Invoking the retriever means the most relevant documents are sorted first
-    sorted_docs = retriever.invoke(query)
-
-    # Define the regex pattern to find the start time
-    # pattern = r'start="([\d/]+ [\d:]+)"'
-    pattern = r'start="([\d/]+ [\d:]+ ?[APM]{0,2})"'
-
-    docs = []
-    dts = []
-
-    # Find the startime of the document
-    for doc in sorted_docs:
-        doc_content = doc.page_content
-        start_dt_match = re.search(pattern, doc_content)
-        if start_dt_match:
-            start_dt = start_dt_match.group(1)
-            parsed = False
-            # Try different date formats
-            for fmt in (
-                "%m/%d/%y %I:%M %p",
-                "%m/%d/%Y %I:%M %p",
-                "%m/%d/%y %H:%M",
-                "%m/%d/%Y %H:%M",
-            ):
-                try:
-                    dts.append(datetime.datetime.strptime(start_dt, fmt))
-                    parsed = True
-                    break
-                except ValueError:
-                    continue
-            if not parsed:
-                print(f"Error parsing date: {start_dt}")
-                continue
-        else:
-            print(f"Start time not found., {doc_content}")
-            dts.append(datetime.datetime.min)
-        docs.append(doc_content)
-
-    final_docs = []
-    current_length = 0
-
-    # Add documents until we exceed the allocated context length
-    for i in range(len(docs)):
-        doc_content = docs[i]
-        doc_length = len(tokenizer.encode(doc_content))
-        final_docs.append((dts[i], doc_content))
-        current_length += doc_length
-        if current_length > target_length:
-            break
-
-    # Sort final_docs chronologically
-    final_docs.sort(key=lambda x: x[0])
-
-    # Extract only the document content for the final output
-    final_docs_content = [doc_content for _, doc_content in final_docs]
-
-    return final_docs_content
-
-
-
 def pack_and_trim_prompts(
     instructions: Dict[int, Dict[str, str]],
     ehrs: Dict[int, str],
-    prompt_template: langchain.prompts.PromptTemplate,
+    prompt_string: str,
     context_length: int,
     generation_length: int,
     tokenizer: Any,
-    use_RAG: bool = True,
     verbose: bool = False,
     include_ehr: bool = True,
 ) -> Dict[int, str]:
@@ -276,26 +178,15 @@ def pack_and_trim_prompts(
         patient_id = int(instructions[instruction_id]["patient_id"])
         relevant_ehr = ehrs[patient_id]
 
-        # Calculate how many tokens of EHR we can include in the prompt
         num_tokens_instruction = len(tokenizer.encode(instruction))
-        num_tokens_prompt_template = len(tokenizer.encode(prompt_template.template))
+        num_tokens_prompt_template = len(tokenizer.encode(prompt_string))
         if include_ehr:
             target_ehr_length = context_length - generation_length - num_tokens_prompt_template - num_tokens_instruction
         else:
             target_ehr_length = 0
         if target_ehr_length <= 0:
-            prompt_with_truncated_ehr = prompt_template.format(question=instruction, ehr="")
+            prompt_with_truncated_ehr = prompt_string.format(question=instruction, ehr="")
         else:
-            if use_RAG:
-                # Return a list of the most relevant visit strings
-                most_relevant_visits = retrieve_most_relevant_visits(
-                    ehr_visit_strs=relevant_ehr,
-                    query=instruction,
-                    target_length=target_ehr_length,
-                    tokenizer=tokenizer,
-                )
-                relevant_ehr = "\n".join(most_relevant_visits)
-
             # Do a first pass with a fast tokenizer
             fast_tokenizer = tiktoken.get_encoding("cl100k_base")
             fast_encoded = fast_tokenizer.encode(relevant_ehr)
@@ -307,13 +198,17 @@ def pack_and_trim_prompts(
                 encoded_ehr = tokenizer.encode(fast_truncated_ehr)
                 truncated_encoded_ehr = encoded_ehr[-target_ehr_length:]
                 truncated_ehr = tokenizer.decode(truncated_encoded_ehr)
-                prompt_with_truncated_ehr = prompt_template.format(question=instruction, ehr=truncated_ehr)
+                prompt_with_truncated_ehr = prompt_string.format(question=instruction, ehr=truncated_ehr)
+            else:
+                # If the fast encoding is still too long, just use the full EHR up to allowed length
+                truncated_ehr = fast_tokenizer.decode(fast_encoded[-target_ehr_length:])
+                prompt_with_truncated_ehr = prompt_string.format(question=instruction, ehr=truncated_ehr)
 
-                prompts_map[instruction_id] = prompt_with_truncated_ehr
+        prompts_map[instruction_id] = prompt_with_truncated_ehr
 
-                if verbose:
-                    print(prompt_with_truncated_ehr)
-                    print("~" * 20)
+        if verbose:
+            print(prompt_with_truncated_ehr)
+            print("~" * 20)
     return prompts_map
 
 
@@ -322,7 +217,6 @@ def preprocess_prompts(
     generation_length,
     path_to_instructions,
     path_to_ehrs,
-    use_RAG,
     include_ehr,
     tokenizer,
     codes_only=False,
@@ -347,16 +241,18 @@ def preprocess_prompts(
 
     # CONSTRUCT & TRUNCATE PROMPTS #
     print("Constructing prompts using instructions and EHRs...")
-    prompt_string="Instruction: Answer the following question based on the EHR:\n\nEHR: {ehr}\n\nQuestion: {question}\n\nAnswer:"
-    prompt_template = langchain.prompts.PromptTemplate.from_template(prompt_string)
+    prompt_string = (
+        "Instruction: Answer the following question based on the EHR:\n\n"
+        "EHR: {ehr}\n\nQuestion: {question}\n\nAnswer:"
+    )
+
     filled_prompts = pack_and_trim_prompts(
         instructions=instructions,
         ehrs=ehrs,
-        prompt_template=prompt_template,
+        prompt_string=prompt_string,
         context_length=target_context_length,
         generation_length=generation_length,
         tokenizer=tokenizer,
-        use_RAG=use_RAG,
         verbose=False,
         include_ehr=include_ehr,
     )
@@ -415,7 +311,6 @@ def return_dataset_dataframe(max_length: int, data_path: str) -> pd.DataFrame:
     path_to_ehrs = os.path.join(data_path, "medalign_ehr_xml")
     path_to_reference_responses = os.path.join(data_path, "clinician-instruction-responses.tsv")
     check_file_exists(path_to_reference_responses, msg=f"[MedAlignScenario] Required clinician responses file not found: '{path_to_reference_responses}'")
-    use_RAG = False
     include_ehr = True
     tokenizer = "tiktoken"
 
@@ -424,7 +319,6 @@ def return_dataset_dataframe(max_length: int, data_path: str) -> pd.DataFrame:
         generation_length=generation_length,
         path_to_instructions=path_to_instructions,
         path_to_ehrs=path_to_ehrs,
-        use_RAG=use_RAG,
         include_ehr=include_ehr,
         tokenizer=tokenizer,
     )

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"