MultiMedQA (#1198)

* multimedqa

* Update medqa.yaml

* move to benchmarks folder

* add README.md

---------
Co-authored-by: Lintang Sutawika <lintang@sutawika.com>

lm_eval/tasks/benchmarks/multimedqa/README.md

+# MultiMedQA (multiple-choice subset)
+
+### Paper
+
+Title: Large Language Models Encode Clinical Knowledge
+
+Abstract: https://arxiv.org/abs/2212.13138
+
+A benchmark combining four existing multiple-choice question answering datasets spanning professional medical exams and research queries.
+
+### Citation
+
+```
+@Article{Singhal2023,
+author={Singhal, Karan and Azizi, Shekoofeh and Tu, Tao and Mahdavi, S. Sara and Wei, Jason and Chung, Hyung Won and Scales, Nathan and Tanwani, Ajay and Cole-Lewis, Heather and Pfohl, Stephen and Payne, Perry and Seneviratne, Martin and Gamble, Paul and Kelly, Chris and Babiker, Abubakr and Sch{\"a}rli, Nathanael and Chowdhery, Aakanksha and Mansfield, Philip and Demner-Fushman, Dina and Ag{\"u}era y Arcas, Blaise and Webster, Dale and Corrado, Greg S. and Matias, Yossi and Chou, Katherine and Gottweis, Juraj and Tomasev, Nenad and Liu, Yun and Rajkomar, Alvin and Barral, Joelle and Semturs, Christopher and Karthikesalingam, Alan and Natarajan, Vivek},
+title={Large language models encode clinical knowledge},
+journal={Nature},
+year={2023},
+month={Aug},
+day={01},
+volume={620},
+number={7972},
+pages={172-180},
+issn={1476-4687},
+doi={10.1038/s41586-023-06291-2},
+url={https://doi.org/10.1038/s41586-023-06291-2}
+}
+```
+
+### Tasks
+
+* [PubMedQA](https://pubmedqa.github.io/) - 1,000 expert-labeled Q&A pairs where a question and corresponding PubMed abstract as context is given and the a yes/maybe/no answer must be produced. Unlike the rest of the tasks in this suite, PubMedQA is a closed-domain Q&A task.
+* [MedQA](https://github.com/jind11/MedQA) - US Medical License Exam (USMLE) questions with 4 or 5 possible answers. Typically, only the 4-option questions are used.
+* [MedMCQA](https://medmcqa.github.io/) - 4-option multiple choice questions from Indian medical entrance examinations, >191k total questions.
+* [MMLU](https://arxiv.org/abs/2009.03300) - 4-option multiple choice exam questions from a variety of domains. The following 6 domains are utilized here:
+	* Anatomy
+	* Clinical Knowledge
+	* College Medicine
+	* Medical Genetics
+	* Professional Medicine
+	* College Biology
+
+Note that MultiMedQA also includes some short-form and long-form Q&A tasks (LiveQA, MedicationQA, HealthSearchQA). Evaluation on these tasks is usually done by experts and is not typically performed automatically, and therefore is ignored here.

lm_eval/tasks/benchmarks/multimedqa/multimedqa.yaml

+group: multimedqa
+task:
+  - pubmedqa
+  - medmcqa
+  - medqa_4options
+  - mmlu_anatomy
+  - mmlu_clinical_knowledge
+  - mmlu_college_medicine
+  - mmlu_medical_genetics
+  - mmlu_professional_medicine
+  - mmlu_college_biology

lm_eval/tasks/medmcqa/medmcqa.yaml

+task: medmcqa
+dataset_path: medmcqa
+output_type: multiple_choice
+training_split: train
+validation_split: validation
+test_split: validation
+doc_to_text: !function utils_medmcqa.doc_to_text
+doc_to_target: cop
+doc_to_choice: [ 'A','B','C','D' ]
+should_decontaminate: true
+doc_to_decontamination_query: "{{question}}"
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+  - metric: acc_norm
+    aggregation: mean
+    higher_is_better: true

lm_eval/tasks/medmcqa/utils_medmcqa.py

+# Copied from Master
+def doc_to_text(doc) -> str:
+    """
+    Question: <question>
+    Choices:
+    A. <choice1>
+    B. <choice2>
+    C. <choice3>
+    D. <choice4>
+    Answer:
+    """
+    choices = [doc["opa"], doc["opb"], doc["opc"], doc["opd"]]
+    option_choices = {'A': choices[0], 'B': choices[1], 'C': choices[2], 'D': choices[3]}
+
+    prompt = "Question: " + doc["question"] + "\nChoices:\n"
+    for choice, option in option_choices.items():
+        prompt += f"{choice.upper()}. {option}\n"
+    prompt += "Answer:"
+    return prompt

lm_eval/tasks/medqa/medqa.yaml

+task: medqa_4options
+dataset_path: GBaker/MedQA-USMLE-4-options-hf
+output_type: multiple_choice
+training_split: train
+validation_split: validation
+test_split: test
+doc_to_text: !function preprocess_medqa.doc_to_text
+doc_to_target: !function preprocess_medqa.doc_to_target
+doc_to_choice: [ 'A', 'B', 'C', 'D' ]
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+  - metric: acc_norm
+    aggregation: mean
+    higher_is_better: true

lm_eval/tasks/medqa/preprocess_medqa.py

+def doc_to_text(doc) -> str:
+    option_choices = {'A': doc["ending0"], 'B': doc["ending1"], 'C': doc["ending2"], 'D': doc["ending3"]}
+    answers = "".join((f"{k}. {v}\n") for k, v in option_choices.items())
+    return f"Question: {doc['sent1']}\n{answers}Answer:"
+
+
+def doc_to_target(doc) -> int:
+    return doc["label"]