Merge branch 'upstream' into 'mmlu-pro'

add tokenizer logs info (#1731)

See merge request shijie.yu/lm-evaluation-harness!4

lm_eval/tasks/leaderboard/mmlu_pro/utils.py

+import string
+
+
+def doc_to_text(doc):
+    doc_to_text = f"{doc['question']}\n"
+
+    for i in range(len(doc["options"])):
+        doc_to_text += f"{string.ascii_uppercase[i]}. {doc['options'][i]}\n"
+
+    doc_to_text += "Answer:"
+    return doc_to_text
+
+
+def doc_to_choice(doc):
+    return [string.ascii_uppercase[i] for i in range(len(doc["options"]))]

lm_eval/tasks/leaderboard/musr/_musr.yaml

+group: leaderboard_musr
+task:
+  - leaderboard_musr_murder_mysteries
+  - leaderboard_musr_object_placements
+  - leaderboard_musr_team_allocation

lm_eval/tasks/leaderboard/musr/_template_yaml

+dataset_path: TAUR-Lab/MuSR
+output_type: multiple_choice
+doc_to_text: !function utils.doc_to_text
+doc_to_target: "{{answer_choice}}"
+doc_to_choice: "{{choices}}"
+metric_list:
+  - metric: acc_norm
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/leaderboard/musr/musr_murder_mysteries.yaml

+include: "_template_yaml"
+task: leaderboard_musr_murder_mysteries
+test_split: murder_mysteries

lm_eval/tasks/leaderboard/musr/musr_object_placements.yaml

+include: "_template_yaml"
+task: leaderboard_musr_object_placements
+test_split: object_placements

lm_eval/tasks/leaderboard/musr/musr_team_allocation.yaml

+include: "_template_yaml"
+task: leaderboard_musr_team_allocation
+test_split: team_allocation

lm_eval/tasks/leaderboard/musr/utils.py

+import ast
+
+
+def doc_to_choice(doc):
+    """
+    Convert a doc to a choice.
+    """
+    return ast.literal_eval(doc["choices"])
+
+
+DOC_TO_TEXT = "{narrative}\n\n" "{question}\n\n" "{choices}\n" "Answer:"
+
+
+def doc_to_text(doc):
+    """
+    Convert a doc to text.
+    """
+    choices = ""
+    for i, choice in enumerate(ast.literal_eval(doc["choices"])):
+        choices += f"{i+1} - {choice}\n"
+
+    text = DOC_TO_TEXT.format(
+        narrative=doc["narrative"], question=doc["question"], choices=choices
+    )
+
+    return text

lm_eval/tasks/mathqa/mathqa.yaml

-group:
+tag:
   - math_word_problems
 task: mathqa
 dataset_path: math_qa

lm_eval/tasks/med_concepts_qa/README.md

+# MedConceptsQA
+
+### Paper
+
+Title: `MedConceptsQA: Open Source Medical Concepts QA Benchmark`
+
+Abstract: https://arxiv.org/abs/2405.07348
+
+MedConceptsQA is a dedicated open source benchmark for medical concepts question answering. The benchmark comprises of questions of various medical concepts across different vocabularies: diagnoses, procedures, and drugs.
+
+The questions are categorized into three levels of difficulty: easy, medium, and hard.
+
+Our benchmark serves as a valuable resource for evaluating the
+abilities of Large Language Models to interpret medical codes and distinguish
+between medical concepts.
+
+### Citation
+
+```
+@article{shoham2024medconceptsqa,
+  title={MedConceptsQA--Open Source Medical Concepts QA Benchmark},
+  author={Shoham, Ofir Ben and Rappoport, Nadav},
+  journal={arXiv preprint arXiv:2405.07348},
+  year={2024}
+}
+```
+
+### Groups and Tasks
+
+#### Groups
+
+* `med_concepts_qa`: Contains all the QA tasks (diagnosis, procedures ,and drugs).
+
+#### Tasks
+
+
+* `med_concepts_qa_icd9cm` - ICD9-CM (diagnosis codes, ICD9 format) question-answering. This involves providing information, clarifications, and answering questions related to ICD-9-CM (International Classification of Diseases, 9th Revision, Clinical Modification) diagnosis codes.
+
+
+* `med_concepts_qa_icd10cm` - ICD10-CM (diagnosis codes, ICD10 format) question-answering. This involves providing information, clarifications, and answering questions related to ICD-10-CM (International Classification of Diseases, 10th Revision, Clinical Modification) diagnosis codes.
+
+
+* `med_concepts_qa_icd9proc` - ICD9-Proc (procedure codes, ICD9 format) question-answering. This involves providing information, clarifications, and answering questions related to ICD-9-PCS (International Classification of Diseases, 9th Revision, Procedure Coding System) procedure codes.
+
+
+* `med_concepts_qa_icd10proc` - ICD10-Proc (procedure codes, ICD10 format) question-answering. This involves providing information, clarifications, and answering questions related to ICD-10-PCS (International Classification of Diseases, 10th Revision, Procedure Coding System) procedure codes.
+
+
+* `med_concepts_qa_atc` - ATC (Anatomical Therapeutic Chemical Classification System) question-answering. This involves providing information, clarifications, and answering questions related to the ATC classification system, which is used for the classification of drugs and other medical products according to the organ or system on which they act and their therapeutic, pharmacological, and chemical properties.

lm_eval/tasks/med_concepts_qa/_default_template_yaml

+dataset_path: ofir408/MedConceptsQA
+output_type: multiple_choice
+description: "Answer A,B,C,D according to the answer to this multiple choice question.\n"
+fewshot_split: dev
+fewshot_config:
+  sampler: first_n
+num_fewshot: 4
+test_split: test
+doc_to_text: "{{question}}\nAnswer:"
+doc_to_target: answer_id
+doc_to_choice: ['A', 'B', 'C', 'D']
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true

lm_eval/tasks/med_concepts_qa/_generate_configs.py

+from typing import List
+
+import yaml
+
+
+def generate_yaml_content(vocab_name: str, level: str):
+    content = {
+        "dataset_name": f"{vocab_name}_{level}",
+        "tag": f"med_concepts_qa_{vocab_name}_tasks",
+        "include": "_default_template_yaml",
+        "task": f"med_concepts_qa_{vocab_name}_{level}",
+        "task_alias": f"{vocab_name}_{level}",
+    }
+    return content
+
+
+def generate_yaml_files(
+    vocab_names: List[str], levels: List[str], file_name_prefix: str
+):
+    for vocab_name in vocab_names:
+        for level in levels:
+            yaml_content = generate_yaml_content(vocab_name, level)
+            filename = f"{file_name_prefix}_{vocab_name}_{level}.yaml"
+            with open(filename, "w") as yaml_file:
+                yaml.dump(yaml_content, yaml_file, default_flow_style=False)
+            print(f"Done to generated {filename}")
+
+
+if __name__ == "__main__":
+    generate_yaml_files(
+        vocab_names=["icd9cm", "icd10cm", "icd9proc", "icd10proc", "atc"],
+        levels=["easy", "medium", "hard"],
+        file_name_prefix="med_concepts_qa",
+    )

lm_eval/tasks/med_concepts_qa/_med_concepts_qa.yaml

+group: med_concepts_qa
+task:
+  - med_concepts_qa_icd9cm
+  - med_concepts_qa_icd10cm
+  - med_concepts_qa_icd9proc
+  - med_concepts_qa_icd10proc
+  - med_concepts_qa_atc
+aggregate_metric_list:
+  - metric: acc
+    aggregation: mean

lm_eval/tasks/med_concepts_qa/_med_concepts_qa_atc.yaml

+group: med_concepts_qa_atc
+task:
+  - med_concepts_qa_atc_tasks
+aggregate_metric_list:
+  - metric: acc
+    aggregation: mean

lm_eval/tasks/med_concepts_qa/_med_concepts_qa_icd10cm.yaml

+group: med_concepts_qa_icd10cm
+task:
+  - med_concepts_qa_icd10cm_tasks
+aggregate_metric_list:
+  - metric: acc
+    aggregation: mean

lm_eval/tasks/med_concepts_qa/_med_concepts_qa_icd10proc.yaml

+group: med_concepts_qa_icd10proc
+task:
+  - med_concepts_qa_icd10proc_tasks
+aggregate_metric_list:
+  - metric: acc
+    aggregation: mean

lm_eval/tasks/med_concepts_qa/_med_concepts_qa_icd9cm.yaml

+group: med_concepts_qa_icd9cm
+task:
+  - med_concepts_qa_icd9cm_tasks
+aggregate_metric_list:
+  - metric: acc
+    aggregation: mean

lm_eval/tasks/med_concepts_qa/_med_concepts_qa_icd9proc.yaml

+group: med_concepts_qa_icd9proc
+task:
+  - med_concepts_qa_icd9proc_tasks
+aggregate_metric_list:
+  - metric: acc
+    aggregation: mean

lm_eval/tasks/med_concepts_qa/med_concepts_qa_atc_easy.yaml

+dataset_name: atc_easy
+include: _default_template_yaml
+tag: med_concepts_qa_atc_tasks
+task: med_concepts_qa_atc_easy
+task_alias: atc_easy

lm_eval/tasks/med_concepts_qa/med_concepts_qa_atc_hard.yaml

+dataset_name: atc_hard
+include: _default_template_yaml
+tag: med_concepts_qa_atc_tasks
+task: med_concepts_qa_atc_hard
+task_alias: atc_hard

lm_eval/tasks/med_concepts_qa/med_concepts_qa_atc_medium.yaml

+dataset_name: atc_medium
+include: _default_template_yaml
+tag: med_concepts_qa_atc_tasks
+task: med_concepts_qa_atc_medium
+task_alias: atc_medium