from dataclasses import dataclass from enum import Enum @dataclass class Task: benchmark: str metric: str col_name: str # Select your tasks here # --------------------------------------------------- class Tasks(Enum): # task_key in the json file, metric_key in the json file, name to display in the leaderboard task0 = Task("medmcqa", "acc,none", "MedMCQA") task1 = Task("medqa_4options", "acc,none", "MedQA") task2 = Task("mmlu_anatomy", "acc,none", "MMLU Anatomy") task3 = Task("mmlu_clinical_knowledge", "acc,none", "MMLU Clinical Knowledge") task4 = Task("mmlu_college_biology", "acc,none", "MMLU College Biology") task5 = Task("mmlu_college_medicine", "acc,none", "MMLU College Medicine") task6 = Task("mmlu_medical_genetics", "acc,none", "MMLU Medical Genetics") task7 = Task("mmlu_professional_medicine", "acc,none", "MMLU Professional Medicine") task8 = Task("pubmedqa", "acc,none", "PubMedQA") NUM_FEWSHOT = 0 # Change with your few shot # --------------------------------------------------- TITLE = """
Descriptive Alt Text

Open Medical-LLM Leaderboard

""" # What does your leaderboard evaluate? INTRODUCTION_TEXT = """ 🩺 The Open Medical LLM Leaderboard aims to track, rank and evaluate the performance of large language models (LLMs) on medical question answering tasks. It evaluates LLMs across a diverse array of medical datasets, including MedQA (USMLE), PubMedQA, MedMCQA, and subsets of MMLU related to medicine and biology. The leaderboard offers a comprehensive assessment of each model's medical knowledge and question answering capabilities. The datasets cover various aspects of medicine such as general medical knowledge, clinical knowledge, anatomy, genetics, and more. They contain multiple-choice and open-ended questions that require medical reasoning and understanding. More details on the datasets can be found in the "LLM Benchmarks Details" section below. The main evaluation metric used is Accuracy (ACC). Submit a model for automated evaluation on the "Submit" page. If you have comments or suggestions on additional medical datasets to include, please reach out to us in our discussion forum. Evaluation Purpose: The primary role of this leaderboard is to assess and compare the performance of the models. It does not facilitate the distribution, deployment, or clinical use of these models. The models on this leaderboard are not approved for clinical use and are intended for research purposes only. Please refer to the "Advisory Notice" section in the "About" page. The backend of the Open Medical LLM Leaderboard uses the Eleuther AI Language Model Evaluation Harness. More technical details can be found in the "About" page. The GPT-4, and Med-PaLM-2 results are taken from their official papers. Since Med-PaLM doesn't provide zero-shot accuracy, we are using 5-shot accuracy from their paper for comparison. All results presented are in the zero-shot setting, except for Med-PaLM-2 which use 5-shot accuracy. Gemini results are taken from recent Clinical-NLP (NAACL 24) Paper """ LLM_BENCHMARKS_TEXT = f"""

Why Leaderboard?

Evaluating the medical knowledge and clinical reasoning capabilities of LLMs is crucial as they are increasingly being applied to healthcare and biomedical applications. The Open Medical LLM Leaderboard provides a platform to assess the latest LLMs on their performance on a variety of medical question answering tasks. This can help identify the strengths and gaps in medical understanding of current models.

How it works

📈 We evaluate the models on 9 medical Q&A datasets using the Eleuther AI Language Model Evaluation Harness , a unified framework to test language models on different tasks.

Advisory Notice

The Open Medical-LLM Leaderboard showcases medical models intended solely for research and development purposes. It is important to be aware of the following: Evaluation Purpose: The primary role of this leaderboard is to assess and compare the performance of the models. It does not facilitate the distribution, deployment, or clinical use of these models. Regulatory Status: The models listed on this leaderboard have not been approved or registered by any regulatory authorities, including the US FDA, the European Medicines Agency (EMA), Health Canada, or the Therapeutic Goods Administration (TGA) in Australia. They are not listed in the US FDA Database for approved AI in healthcare or the EUDAMED database. Disclaimer: These models are not intended for direct patient care, clinical decision support, or any other professional medical purposes. Their use should be limited to research, development, and exploratory applications by qualified individuals who understand their limitations and the regulatory requirements. Risk Warning: The outputs of these models may contain inaccuracies, biases, or misalignments that could pose risks if relied upon for medical decision-making. The models' performance has not been rigorously evaluated in randomized controlled trials or real-world healthcare environments. Research Tool Only: The models on this leaderboard are intended solely as research tools to assist healthcare professionals and should never be considered a replacement for the professional judgment and expertise of a qualified medical doctor. Further Validation Needed: Proper adaptation and validation of these models for specific medical use cases would require significant additional work, some of them including: 1) Thorough testing and evaluation in relevant clinical scenarios. 2) Alignment with evidence-based guidelines and best practices. 3) Mitigation of potential biases and failure modes. 4) Integration with human oversight and interpretation. 5) Compliance with regulatory and ethical standards. Always consult a qualified healthcare provider for personal medical needs.

About Open Life Science AI

An Open Life Science Project to Benchmark and Track AI Progress, Share Models and Datasets in the Life Science Field. More info

Datasets

Evaluation Metric

Metric Accuracy (ACC) is used as the main evaluation metric across all datasets.

Details and Logs

Detailed results are available in the results directory:

https://huggingface.co/datasets/openlifescienceai/results

Input/outputs for each model can be found in the details page accessible by clicking the 📄 emoji next to the model name.

Reproducibility

To reproduce the results, you can run this evaluation script:

python eval_medical_llm.py

To evaluate a specific dataset on a model, use the EleutherAI LLM Evaluation Harness:

python main.py --model=hf-auto --model_args="pretrained=<model>,revision=<revision>,parallelize=True" --tasks=<dataset> --num_fewshot=<n_shots> --batch_size=1 --output_path=<output_dir>

Note some datasets may require additional setup, refer to the Evaluation Harness documentation.

Adjust batch size based on your GPU memory if not using parallelism. Minor variations in results are expected with different batch sizes due to padding.

Icons

Missing icons indicate the model info is not yet added, feel free to open an issue to include it!

""" LLM_BENCHMARKS_DETAILS = f""" Datasets MedQA (USMLE) - 1273 real-world questions from the US Medical License Exams (USMLE) to test general medical knowledge PubMedQA - 500 questions constructed from PubMed article titles along with the abstracts as context to test understanding of biomedical research MedMCQA - 4183 questions from Indian medical entrance exams (AIIMS & NEET PG) spanning 2.4k healthcare topics MMLU-Clinical knowledge - 265 multiple choice questions on clinical knowledge MMLU-Medical genetics - 100 MCQs on medical genetics MMLU-Anatomy - 135 anatomy MCQs MMLU-Professional medicine - 272 MCQs on professional medicine MMLU-College biology - 144 MCQs on college-level biology MMLU-College medicine - 173 college medicine MCQs Metric Accuracy (ACC) is used as the main evaluation metric across all datasets Details and logs Detailed results are available in the results directory: https://huggingface.co/spaces/openlifescienceai/open_medical_llm_leaderboard/tree/main/results Input/outputs for each model can be found in the details page accessible by clicking the 📄 emoji next to the model name Reproducibility To reproduce the results, you can run this evaluation script: python eval_medical_llm.py. To evaluate a specific dataset on a model, use the EleutherAI LLM Evaluation Harness: python main.py --model=hf-auto --model_args="pretrained=,revision=,parallelize=True" --tasks= --num_fewshot= --batch_size=1 --output_path= Note some datasets may require additional setup, refer to the Evaluation Harness documentation. Adjust batch size based on your GPU memory if not using parallelism. Minor variations in results are expected with different batch sizes due to padding. Icons 🟢 Pre-trained model 🔶 Fine-tuned model ? Unknown model type ⭕ instruction-tuned 🟦 RL-tuned Missing icons indicate the model info is not yet added, feel free to open an issue to include it! """ Advisory_Notice = """The Open Medical-LLM Leaderboard showcases medical models intended solely for research and development purposes. It is important to be aware of the following: Regulatory Status: The models listed on this leaderboard have not been approved or registered by any regulatory authorities, including the US FDA, the European Medicines Agency (EMA), Health Canada, or the Therapeutic Goods Administration (TGA) in Australia. They are not listed in the US FDA Database for approved AI in healthcare or the EUDAMED database. As such, they are not compliant with regulations such as US FDA 21 CFR 820 and EU MDR 2017/745. Disclaimer: These models are not intended for direct patient care, clinical decision support, or any other professional medical purposes. Their use should be limited to research, development, and exploratory applications by qualified individuals who understand their limitations and the regulatory requirements. Risk Warning: The outputs of these models may contain inaccuracies, biases, or misalignments that could pose risks if relied upon for medical decision-making. The models' performance has not been rigorously evaluated in randomized controlled trials or real-world healthcare environments. Research Tool Only: The models on this leaderboard are intended solely as research tools to assist healthcare professionals and should never be considered a replacement for the professional judgment and expertise of a qualified medical doctor. Further Validation Needed: Proper adaptation and validation of these models for specific medical use cases would require significant additional work, including: 1) Thorough testing and evaluation in relevant clinical scenarios. 2) Alignment with evidence-based guidelines and best practices. 3) Mitigation of potential biases and failure modes. 4) Integration with human oversight and interpretation. 5) Compliance with regulatory and ethical standards. For any legal inquiries or concerns, please contact the authors of the MedPaLM papers directly. Always consult a qualified healthcare provider for personal medical needs.""" FAQ_TEXT = """ FAQ 1) Submitting a model XXX 2) Model results XXX 3) Editing a submission XXX """ EVALUATION_QUEUE_TEXT = """ Evaluation Queue for the Open Medical LLM Leaderboard Models added here will be automatically evaluated. Before submitting a model 1) Verify loading with AutoClasses: from transformers import AutoConfig, AutoModel, AutoTokenizer config = AutoConfig.from_pretrained("model-name", revision=revision) model = AutoModel.from_pretrained("model-name", revision=revision) tokenizer = AutoTokenizer.from_pretrained("model-name", revision=revision) Debug any loading errors before submission. Make sure the model is public. Note: Models that require use_remote_code=True are not yet supported. 2) Convert weights to safetensors This allows faster loading and enables showing model parameters in the Extended Viewer. 3) Select correct precision Incorrect precision (e.g. loading bf16 as fp16) can cause NaN errors for some models. Debugging failing models For models in FAILED status, first ensure the above checks are done. Then test running the Eleuther AI Harness locally using the command in the "Reproducibility" section, specifying all arguments. Add --limit to evaluate on fewer examples per task. """ CITATION_BUTTON_LABEL = "Copy the citation snippet" CITATION_BUTTON_TEXT = r""" @misc{openlifescienceai/open_medical_llm_leaderboard, author = {Ankit Pal and Pasquale Minervini and Andreas Geert Motzfeldt and Beatrice Alex}, title = {openlifescienceai/open_medical_llm_leaderboard}, year = {2024}, publisher = {Hugging Face}, howpublished = "\url{https://huggingface.co/spaces/openlifescienceai/open_medical_llm_leaderboard}" } @misc{singhal2022large, title={Large Language Models Encode Clinical Knowledge}, author={Karan Singhal et al.}, year={2022}, eprint={2212.13138}, archivePrefix={arXiv}, primaryClass={cs.CL} } """