crfm-helm 0.5.4__py3-none-any.whl → 0.5.5__py3-none-any.whl

helm/benchmark/scenarios/scenario.py

@@ -1,6 +1,6 @@
 from abc import ABC, abstractmethod
 from dataclasses import dataclass, field, replace
-from typing import List, Optional, Tuple
+from typing import Dict, List, Optional, Tuple, Any
 import os
 from pathlib import PurePath
 import inspect
@@ -67,6 +67,11 @@ class Input:
     multimedia_content: Optional[MultimediaObject] = None
     """A single input can consists of multimodal content interleaved (e.g., text, image, text, ...)."""
 
+    messages: Optional[List[Dict[str, str]]] = None
+    """Used for chat models.
+    If messages is specified for a chat model, the prompt is ignored.
+    Otherwise, the client should convert the prompt into a message."""
+
 
 @dataclass(frozen=True)
 class PassageQuestionInput(Input):
@@ -153,6 +158,9 @@ class Instance:
     contrast_references: Optional[List[List[Reference]]] = None
     """References for the perturbed input above (if available)"""
 
+    extra_data: Optional[Dict[str, Any]] = None
+    """Extra data required by the scenario e.g. chain-of-thought annotations"""
+
     @property
     def first_correct_reference(self) -> Optional[Reference]:
         """Return the first correct reference."""

helm/benchmark/scenarios/{bhasa_scenario.py → seahelm_scenario.py}

@@ -19,7 +19,7 @@ from helm.benchmark.scenarios.scenario import (
 from helm.common.general import ensure_file_downloaded
 from helm.common.hierarchical_logger import hlog
 
-# BHASA Scenarios
+# SEA-HELM Scenarios
 #   A. Natural Language Understanding
 #   B. Natural Language Generation
 #   C. Natural Language Reasoning
@@ -95,7 +95,12 @@ class TyDiQAScenario(Scenario):
         self.splits = {"train": TRAIN_SPLIT, "validation": TEST_SPLIT}
 
     def get_instances(self, output_path) -> List[Instance]:
-        dataset = datasets.load_dataset("khalidalt/tydiqa-goldp", "indonesian")
+        dataset = datasets.load_dataset(
+            "khalidalt/tydiqa-goldp",
+            "indonesian",
+            revision="7d69b53c9c8187ae7e21d8441362efa1a7e3013d",
+            trust_remote_code=True,
+        )
 
         outputs = []
         for split in self.splits.keys():

helm/benchmark/scenarios/self_instruct_scenario.py

@@ -3,7 +3,7 @@ import os
 from typing import List
 
 from helm.common.general import ensure_file_downloaded
-from .scenario import CORRECT_TAG, Reference, Scenario, Instance, Input, TEST_SPLIT, Output
+from helm.benchmark.scenarios.scenario import CORRECT_TAG, Reference, Scenario, Instance, Input, TEST_SPLIT, Output
 
 
 class SelfInstructScenario(Scenario):

helm/benchmark/scenarios/shc_bmt_scenario.py

@@ -0,0 +1,69 @@
+import sys
+import csv
+from typing import Dict, List
+
+from helm.benchmark.scenarios.scenario import (
+    Input,
+    Scenario,
+    Instance,
+    TEST_SPLIT,
+    CORRECT_TAG,
+    Reference,
+    Output,
+)
+
+csv.field_size_limit(sys.maxsize)
+
+
+class SHCBMTMedScenario(Scenario):
+    """
+    This benchmark dataset was built from a patient status gold-standard
+    for specific questions asked after a bone marrow transplant has taken place.
+    """
+
+    name = "shc_bmt_med"
+    description = (
+        "A dataset containing patient notes with associated "
+        "questions and answers related to bone marrow transplantation."
+    )
+    tags = ["knowledge", "reasoning", "biomedical"]
+
+    POSSIBLE_ANSWER_CHOICES: List[str] = ["A", "B"]
+
+    def create_benchmark(self, csv_path) -> Dict[str, str]:
+        data = {}
+        with open(csv_path, "r") as file:
+            reader = csv.DictReader(file)
+            for row in reader:
+                question = row["prompt"]
+                context = row["context"]
+                answer = row["label"]
+                prompt = (
+                    f"Provide an answer to the following {question} with the following context: {context} "
+                    ", Answer the question with a 'A' for yes or 'B' for no. Do not provide any additional "
+                    "details or response, just a simple A or B response."
+                )
+                data[prompt] = answer
+        return data
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        data_path = "/dbfs/mnt/azure_adbfs/Files/medhelm/medhelm-BMT-dataset_filtered.csv"
+
+        instances: List[Instance] = []
+        benchmark_data = self.create_benchmark(data_path)
+
+        for prompt, answer in benchmark_data.items():
+            assert answer in SHCBMTMedScenario.POSSIBLE_ANSWER_CHOICES
+            references: List[Reference] = [
+                Reference(Output(text=pred_answer), tags=[CORRECT_TAG] if pred_answer == answer else [])
+                for pred_answer in SHCBMTMedScenario.POSSIBLE_ANSWER_CHOICES
+            ]
+            instances.append(
+                Instance(
+                    input=Input(text=prompt),
+                    references=references,  # [Reference(Output(text=answer), tags=[CORRECT_TAG])],
+                    split=TEST_SPLIT,
+                )
+            )
+
+        return instances

helm/benchmark/scenarios/shc_cdi_scenario.py

@@ -0,0 +1,70 @@
+import sys
+import csv
+from typing import Dict, List
+
+from helm.benchmark.scenarios.scenario import (
+    Input,
+    Scenario,
+    Instance,
+    TEST_SPLIT,
+    CORRECT_TAG,
+    Reference,
+    Output,
+)
+
+csv.field_size_limit(sys.maxsize)
+
+
+class SHCCDIMedScenario(Scenario):
+    """
+    This benchmark dataset was built from Clinical Document Integrity (CDI)
+    notes were there are verifications of clinical activities. The idea behind
+    it was to assess an LLM capability to answer these questions from previous notes.
+    """
+
+    name = "shc_cdi_med"
+    description = (
+        "A dataset built from Clinical Document Integrity (CDI) notes, to assess "
+        "the ability to answer verification questions from previous notes."
+    )
+    tags = ["knowledge", "reasoning", "biomedical"]
+
+    POSSIBLE_ANSWER_CHOICES: List[str] = ["A", "B"]
+
+    def create_benchmark(self, csv_path) -> Dict[str, str]:
+        data = {}
+        with open(csv_path, "r") as file:
+            reader = csv.DictReader(file)
+            for row in reader:
+                question = row["prompt"]
+                context = row["context"]
+                answer = row["label"]
+                prompt = (
+                    f"Provide an answer to the following {question} with the following context: {context} , "
+                    "Answer the question with either 'A' for yes or 'B' for no. Do not provide any "
+                    "additional details or response, just a simple A or B response."
+                )
+                data[prompt] = answer
+        return data
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        data_path = "/dbfs/mnt/azure_adbfs/Files/medhelm/medhelm-CDI-dataset_filtered.csv"
+
+        instances: List[Instance] = []
+        benchmark_data = self.create_benchmark(data_path)
+
+        for prompt, answer in benchmark_data.items():
+            assert answer in SHCCDIMedScenario.POSSIBLE_ANSWER_CHOICES
+            references: List[Reference] = [
+                Reference(Output(text=pred_answer), tags=[CORRECT_TAG] if pred_answer == answer else [])
+                for pred_answer in SHCCDIMedScenario.POSSIBLE_ANSWER_CHOICES
+            ]
+            instances.append(
+                Instance(
+                    input=Input(text=prompt),
+                    references=references,
+                    split=TEST_SPLIT,
+                )
+            )
+
+        return instances

helm/benchmark/scenarios/shc_conf_scenario.py

@@ -0,0 +1,70 @@
+import sys
+import csv
+from typing import Dict, List
+
+from helm.benchmark.scenarios.scenario import (
+    Input,
+    Scenario,
+    Instance,
+    TEST_SPLIT,
+    CORRECT_TAG,
+    Reference,
+    Output,
+)
+
+csv.field_size_limit(sys.maxsize)
+
+
+class SHCCONFMedScenario(Scenario):
+    """
+    Benchmark derived from extracting confidential information from clinical notes.
+    From Evaluation of a Large Language Model to Identify Confidential Content in
+    Adolescent Encounter Notes published at https://jamanetwork.com/journals/jamapediatrics/fullarticle/2814109
+    """
+
+    name = "shc_conf_med"
+    description = (
+        "A dataset of clinical notes from adolescent patients used to identify sensitive "
+        "protected health information that should be restricted from parental access."
+    )
+    tags = ["knowledge", "reasoning", "biomedical"]
+
+    POSSIBLE_ANSWER_CHOICES: List[str] = ["A", "B"]
+
+    def create_benchmark(self, csv_path) -> Dict[str, str]:
+        data = {}
+        with open(csv_path, "r") as file:
+            reader = csv.DictReader(file)
+            for row in reader:
+                question = row["prompt"]
+                context = row["context"]
+                answer = row["label"]
+                prompt = (
+                    f"Provide an answer to the following {question} with the following context: {context} "
+                    ", Answer the question with a 'A' for yes or 'B' for no. Do not provide any additional "
+                    "details or response, just a simple A or B response."
+                )
+                data[prompt] = answer
+        return data
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        data_path = "/dbfs/mnt/azure_adbfs/Files/medhelm/medhelm-CONF-dataset_filtered.csv"
+
+        instances: List[Instance] = []
+        benchmark_data = self.create_benchmark(data_path)
+
+        for prompt, answer in benchmark_data.items():
+            assert answer in SHCCONFMedScenario.POSSIBLE_ANSWER_CHOICES
+            references: List[Reference] = [
+                Reference(Output(text=pred_answer), tags=[CORRECT_TAG] if pred_answer == answer else [])
+                for pred_answer in SHCCONFMedScenario.POSSIBLE_ANSWER_CHOICES
+            ]
+            instances.append(
+                Instance(
+                    input=Input(text=prompt),
+                    references=references,  # [Reference(Output(text=answer), tags=[CORRECT_TAG])],
+                    split=TEST_SPLIT,
+                )
+            )
+
+        return instances

helm/benchmark/scenarios/shc_ent_scenario.py

@@ -0,0 +1,72 @@
+import sys
+import csv
+from typing import Dict, List
+
+from helm.benchmark.scenarios.scenario import (
+    Input,
+    Scenario,
+    Instance,
+    TEST_SPLIT,
+    CORRECT_TAG,
+    Reference,
+    Output,
+)
+
+csv.field_size_limit(sys.maxsize)
+
+
+class SHCENTMedScenario(Scenario):
+    """
+    This benchmark dataset was built to assess the capabilities "
+    "of an LLM for referral to the Ear, Nose and Throat department.
+    """
+
+    name = "shc_ent_med"
+    description = (
+        "A dataset designed to evaluate performance in "
+        "identifying appropriate patient referrals to Ear, Nose, and Throat specialists."
+    )
+    tags = ["knowledge", "reasoning", "biomedical"]
+
+    POSSIBLE_ANSWER_CHOICES: List[str] = ["A", "B", "C"]
+
+    def create_benchmark(self, csv_path) -> Dict[str, str]:
+        data = {}
+        counter = 1
+        with open(csv_path, "r") as file:
+            reader = csv.DictReader(file)
+            for row in reader:
+                if row["label"] != "":  # skip rows with character/encoding issues - 79
+                    question = row["prompt"]
+                    context = row["context"]
+                    answer = row["label"]
+                    prompt = (
+                        f"{counter} Provide an answer to the following {question} with the following context:"
+                        f" {context} , Answer the question with either 'A' for yes, 'B' for no, or 'C' for no mention."
+                        " Do not provide any additional details or response, just a simple A, B, or C response."
+                    )
+                    data[prompt] = answer
+                    counter = counter + 1
+        return data
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        data_path = "/dbfs/mnt/azure_adbfs/Files/medhelm/medhelm-ENT-dataset_filtered.csv"
+
+        instances: List[Instance] = []
+        benchmark_data = self.create_benchmark(data_path)
+
+        for prompt, answer in benchmark_data.items():
+            assert answer in SHCENTMedScenario.POSSIBLE_ANSWER_CHOICES
+            references: List[Reference] = [
+                Reference(Output(text=pred_answer), tags=[CORRECT_TAG] if pred_answer == answer else [])
+                for pred_answer in SHCENTMedScenario.POSSIBLE_ANSWER_CHOICES
+            ]
+            instances.append(
+                Instance(
+                    input=Input(text=prompt),
+                    references=references,
+                    split=TEST_SPLIT,
+                )
+            )
+
+        return instances

helm/benchmark/scenarios/shc_gip_scenario.py

@@ -0,0 +1,66 @@
+import sys
+import csv
+from typing import Dict, List
+
+from helm.benchmark.scenarios.scenario import (
+    Input,
+    Scenario,
+    Instance,
+    TEST_SPLIT,
+    CORRECT_TAG,
+    Reference,
+    Output,
+)
+
+csv.field_size_limit(sys.maxsize)
+
+
+class SHCGIPMedScenario(Scenario):
+    """
+    This benchmark dataset was built from a patient referral gold-standard set
+    to a specialty clinic to verify the ability of LLMs for patient hospice referral purposes.
+    """
+
+    name = "shc_gip_med"
+    description = "A dataset evaluating performance in identifying appropriate patient referrals to hospice care."
+    tags = ["knowledge", "reasoning", "biomedical"]
+
+    POSSIBLE_ANSWER_CHOICES: List[str] = ["A", "B"]
+
+    def create_benchmark(self, csv_path) -> Dict[str, str]:
+        data = {}
+        with open(csv_path, "r") as file:
+            reader = csv.DictReader(file)
+            for row in reader:
+                question = row["prompt"]
+                context = row["context"]
+                answer = row["label"]
+                prompt = (
+                    f"Provide an answer to the following {question} with the following context: {context} "
+                    ", Answer the question with a 'A' for yes or 'B' for no. Do not provide any additional "
+                    "details or response, just a simple A or B response."
+                )
+                data[prompt] = answer
+        return data
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        data_path = "/dbfs/mnt/azure_adbfs/Files/medhelm/medhelm-GIP-dataset_filtered.csv"
+
+        instances: List[Instance] = []
+        benchmark_data = self.create_benchmark(data_path)
+
+        for prompt, answer in benchmark_data.items():
+            assert answer in SHCGIPMedScenario.POSSIBLE_ANSWER_CHOICES
+            references: List[Reference] = [
+                Reference(Output(text=pred_answer), tags=[CORRECT_TAG] if pred_answer == answer else [])
+                for pred_answer in SHCGIPMedScenario.POSSIBLE_ANSWER_CHOICES
+            ]
+            instances.append(
+                Instance(
+                    input=Input(text=prompt),
+                    references=references,  # [Reference(Output(text=answer), tags=[CORRECT_TAG])],
+                    split=TEST_SPLIT,
+                )
+            )
+
+        return instances

helm/benchmark/scenarios/shc_ptbm_scenario.py

@@ -0,0 +1,76 @@
+import sys
+import csv
+from typing import Dict, List
+
+from helm.benchmark.scenarios.scenario import (
+    Input,
+    Scenario,
+    Instance,
+    TEST_SPLIT,
+    CORRECT_TAG,
+    Reference,
+    Output,
+)
+
+csv.field_size_limit(sys.maxsize)
+
+
+class SHCPTBMMedScenario(Scenario):
+    """
+    This dataset contains clinical notes from primary care visit encounters of
+    children ages 4-6 years old with ADHD seen at Stanford's community-based primary
+    care network, Packard Children's Health Alliance, between 2015-2019. In this classification
+    task, the LLM is tasked with classifying whether the note contains clinician recommendation
+    for parent training in behavior management, which is the first-line evidence-based treatment
+    for young children with ADHD. From publication: https://doi.org/10.1093/jamia/ocae001
+    """
+
+    name = "shc_ptbm_med"
+    description = (
+        "A dataset that classifies whether a clinical note contains a clinician "
+        "recommendation for parent training in behavior management, which is the first-line "
+        "evidence-based treatment for young children with ADHD."
+    )
+    tags = ["knowledge", "reasoning", "biomedical"]
+
+    POSSIBLE_ANSWER_CHOICES: List[str] = ["A", "B"]
+
+    def create_benchmark(self, csv_path) -> Dict[str, str]:
+        data = {}
+        with open(csv_path, "r") as file:
+            reader = csv.DictReader(file)
+            for row in reader:
+                question = row["prompt"]
+                context = row["context"]
+                answer = row["label"]
+                prompt = (
+                    "You are reviewing a clinical note from health records of children with "
+                    "attention deficit hyperactivity disorder (ADHD) and classifying mentions of "
+                    f"behavioral therapy. Provide an answer to the following {question} with the "
+                    f"following context: {context} , Answer the question with a 'A' for yes or 'B' "
+                    "for no. Do not provide any additional details or response, just a simple A or B response."
+                )
+                data[prompt] = answer
+        return data
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        data_path = "/dbfs/mnt/azure_adbfs/Files/medhelm/medhelm-PTBM-dataset_filtered.csv"
+
+        instances: List[Instance] = []
+        benchmark_data = self.create_benchmark(data_path)
+
+        for prompt, answer in benchmark_data.items():
+            assert answer in SHCPTBMMedScenario.POSSIBLE_ANSWER_CHOICES
+            references: List[Reference] = [
+                Reference(Output(text=pred_answer), tags=[CORRECT_TAG] if pred_answer == answer else [])
+                for pred_answer in SHCPTBMMedScenario.POSSIBLE_ANSWER_CHOICES
+            ]
+            instances.append(
+                Instance(
+                    input=Input(text=prompt),
+                    references=references,  # [Reference(Output(text=answer), tags=[CORRECT_TAG])],
+                    split=TEST_SPLIT,
+                )
+            )
+
+        return instances

helm/benchmark/scenarios/shc_sei_scenario.py

@@ -0,0 +1,89 @@
+import sys
+import csv
+from typing import Dict, List
+
+from helm.benchmark.scenarios.scenario import (
+    Input,
+    Scenario,
+    Instance,
+    TEST_SPLIT,
+    CORRECT_TAG,
+    Reference,
+    Output,
+)
+
+csv.field_size_limit(sys.maxsize)
+
+
+class SHCSEIMedScenario(Scenario):
+    """
+    This dataset contains clinical notes from primary care visit encounters
+    (in-person/telehealth and telephone) of children ages 6-11 years old with ADHD
+    seen at Stanford's community-based primary care network, Packard Children's Health Alliance,
+    between 2015-2022. All children in this dataset were prescribed at least once an ADHD
+    medication (stimulants or non-stimulants) by a primary care clinician. In this
+    classification task, the LLM is tasked with classifying whether the note contains
+    documentation of side effect monitoring (recording of absence or presence of
+    medication side effects), as recommended in clinical practice guidelines.
+    From publication: https://doi.org/10.1542/peds.2024-067223
+    """
+
+    name = "shc_sei_med"
+    description = (
+        "A dataset that classifies whether a clinical note contains documentation "
+        "of side effect monitoring (recording of absence or presence of medication "
+        "side effects), as recommended in clinical practice guidelines."
+    )
+    tags = ["knowledge", "reasoning", "biomedical"]
+
+    POSSIBLE_ANSWER_CHOICES: List[str] = ["A", "B"]
+
+    def create_benchmark(self, csv_path) -> Dict[str, str]:
+        data = {}
+        with open(csv_path, "r") as file:
+            reader = csv.DictReader(file)
+            for row in reader:
+                question = row["prompt"]
+                context = row["context"]
+                answer = row["label"]
+                prompt = (
+                    "You are reviewing a clinical note from health records of children "
+                    "with attention deficit hyperactivity disorder (ADHD). Given the following "
+                    "definitions: side Effects Inquiry (SEI): Explicit documentation by the clinician "
+                    "asking about current side effects related to ADHD medications that the child is "
+                    "taking or documentation of specific ADHD medication side effects experienced "
+                    "by the patient. SEI does *not* include future side effects monitoring, "
+                    "such as documentation of potential ADHD medication side effects, including "
+                    "planning to follow patients to monitor side effects, explaining about "
+                    "potential side effects of an ADHD medication. These documentations are not "
+                    "categorized as SEI because they consist of a plan or an explanation about "
+                    "side effects without actual side effect monitoring taking place, and "
+                    "No Side Effects Inquiry (NSEI): No evidence of side effects monitoring. "
+                    f"Provide an answer to the following {question} with the following context: {context} "
+                    ", Answer the question with a 'A' for yes or 'B' for no. Do not provide any additional "
+                    "details or response, just a simple A or B response."
+                )
+                data[prompt] = answer
+        return data
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        data_path = "/dbfs/mnt/azure_adbfs/Files/medhelm/medhelm-SEI-dataset_filtered.csv"
+
+        instances: List[Instance] = []
+        benchmark_data = self.create_benchmark(data_path)
+
+        for prompt, answer in benchmark_data.items():
+            assert answer in SHCSEIMedScenario.POSSIBLE_ANSWER_CHOICES
+            references: List[Reference] = [
+                Reference(Output(text=pred_answer), tags=[CORRECT_TAG] if pred_answer == answer else [])
+                for pred_answer in SHCSEIMedScenario.POSSIBLE_ANSWER_CHOICES
+            ]
+            instances.append(
+                Instance(
+                    input=Input(text=prompt),
+                    references=references,  # [Reference(Output(text=answer), tags=[CORRECT_TAG])],
+                    split=TEST_SPLIT,
+                )
+            )
+
+        return instances

helm/benchmark/scenarios/shc_sequoia_scenario.py

@@ -0,0 +1,69 @@
+import sys
+import csv
+from typing import Dict, List
+
+from helm.benchmark.scenarios.scenario import (
+    Input,
+    Scenario,
+    Instance,
+    TEST_SPLIT,
+    CORRECT_TAG,
+    Reference,
+    Output,
+)
+
+csv.field_size_limit(sys.maxsize)
+
+
+class SHCSequoiaMedScenario(Scenario):
+    """
+    Benchmark derived from manually curated answers to several questions for Sequoia clinic referrals
+    """
+
+    name = "shc_sequoia_med"
+    description = (
+        "A dataset containing manually curated answers to questions regarding patient referrals to the Sequoia clinic."
+    )
+    tags = ["knowledge", "reasoning", "biomedical"]
+
+    POSSIBLE_ANSWER_CHOICES: List[str] = ["A", "B"]
+
+    def create_benchmark(self, csv_path) -> Dict[str, str]:
+        data = {}
+        counter = 1
+        with open(csv_path, "r") as file:
+            reader = csv.DictReader(file)  # , quoting=csv.QUOTE_MINIMAL
+            for row in reader:
+                question = row["question"]
+                context = row["context"]
+                answer = row["label"]
+                prompt = (
+                    f" {counter} Provide an answer to the following {question} with the following context:"
+                    f" {context} , Answer the question with a 'A' for yes or 'B' for no. Do not provide any "
+                    "additional details or response, just a simple A or B response."
+                )
+                data[prompt] = answer
+                counter += 1
+        return data
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        data_path = "/dbfs/mnt/azure_adbfs/Files/medhelm/medhelm-sequoia-dataset_filtered.csv"
+
+        instances: List[Instance] = []
+        benchmark_data = self.create_benchmark(data_path)
+
+        for prompt, answer in benchmark_data.items():
+            assert answer in SHCSequoiaMedScenario.POSSIBLE_ANSWER_CHOICES
+            references: List[Reference] = [
+                Reference(Output(text=pred_answer), tags=[CORRECT_TAG] if pred_answer == answer else [])
+                for pred_answer in SHCSequoiaMedScenario.POSSIBLE_ANSWER_CHOICES
+            ]
+            instances.append(
+                Instance(
+                    input=Input(text=prompt),
+                    references=references,  # [Reference(Output(text=answer), tags=[CORRECT_TAG])],
+                    split=TEST_SPLIT,
+                )
+            )
+
+        return instances

helm/benchmark/scenarios/simple_safety_tests_scenario.py

@@ -1,7 +1,7 @@
 from typing import List
 from datasets import load_dataset
 
-from .scenario import Scenario, Instance, Input, TEST_SPLIT, Reference, Output
+from helm.benchmark.scenarios.scenario import Scenario, Instance, Input, TEST_SPLIT, Reference, Output
 
 
 class SimpleSafetyTestsScenario(Scenario):