Awesome

Japanese Medical Language Model Evaluation Harness

ワンコマンドで実行可能な医療分野に特化したLLMの日英能力評価プログラム.

Leaderboard

w/o shuffle

with shuffle

(*) The training data of MedSwallow is the Japanese-translated MedQA data, which also includes test split.

• prompt : medpalm_five_choice_cot / medpalm_five_choice_cot_ja, all zero-shot
• quantize : True for 70B models, False for 7B models
• metric : Accuracy based on Gestalt distance (relatively robust)
• use_vllm : off
• environment : NVIDIA A100

Setup

pip install -r requirements.txt
cd dataset
git clone https://github.com/jungokasai/IgakuQA.git
cd ..

Set each dataset as follows

dataset/
- IgakuQA/
    - baseline_results
    - data
        - 2018
        ...
        - 2022
- MedQA
    - usmleqa_en.jsonl
    - usmleqa_ja.jsonl
- MedMCQA
    - medmcqa_en.jsonl
    - medmcqa_ja.jsonl
- JMMLU
    - xxx.csv
- ClinicalQA25
    - clinicalqa_en.jsonl
    - clinicalqa_ja.jsonl

Usage

Ex 1.

python eval_bench.py \
--model_path tokyotech-llm/Swallow-70b-instruct-hf \
--peft AIgroup-CVM-utokyohospital/MedSwallow-70b \
--data IgakuQA \
--prompt alpaca_ja \
--lang ja \
--quantize 

Ex2.

python eval_bench.py \
--model_path tokyotech-llm/Swallow-7b-instruct-hf \
--data MedMCQA \
--prompt medpalm_five_choice_cot_ja \
--lang ja
--use_vllm

Ex3.

python eval_bench.py \
--model_path epfl-llm/meditron-7b \
--data IgakuQA2018 \
--prompt meditron_five_choice \
--lang en
--use_vllm

Ex4.

python eval_bench.py \
--model_path BioMistral/BioMistral-7B \
--data IgakuQA2018 \
--prompt medpalm_five_choice_cot \
--lang en
--use_vllm

Test code

python eval_bench.py \
--model_path tokyotech-llm/Swallow-7b-instruct-hf \
--data sample \
--prompt alpaca_med_five_choice_cot_jp \
--lang ja

Recommended models

• epfl-llm/meditron-7b
• BioMistral/BioMistral-7B
• FreedomIntelligence/Apollo-7B (not supported yet)
• tokyotech-llm/Swallow-7b-instruct-hf

parameters

model_path : huggingface model id lang : "ja" or "en"
prompt : See template.py for options. You can also add your own prompt template and use it. use_vllm : True or False num_gpus : Specify when using vllm, defaults to 1. quantize : True or False. Better to quantize when using 70B LLM.
shuffle : Whether to shuffle the choices.
data :

• "sample" ・・・ for code test
• "IgakuQA" (default) ・・・ Removed non-5-choice Q&As due to its format.
  • "IgakuQA20{18,19,20,21,22}"
• "MedQA"
• "MedMCQA

Evaluation and Metrics

Evaluation Datasets

• ClinicalQA25 from Almanac : 25 Open-ended text generation tasks.
• IgakuQA : Japanese National Medical License Exam.
• MedMCQA : Multi-Subject Multi-Choice Dataset for Medical domain, we only use evaluation split.
• MedQA : Americal National Medical License Exam, we only use evaluation split.

Japanese version of MedMCQA and MedQA were provided at JMedBench by Junfeng Jiang.

For Multiple-choices question-answering

When the choices are
a.) hoge
b.) fuga
...,
the response of the LLM is meant to be "fuga" rather than "b". This can be controlled via prompting to a certain extent.

• accuracy based on exact match
• accuracy based on gestalt match

Notes

• Swallow-7b-instruct-hf, NVIDIA A10G x 1 => 20GB VRAM on GPU, 10 seconds/inference.
• Meditron-7b, NVIDIA A10G x 1 => 20GB VRAM on GPU, 3 minutes/inference.
• greedy sampling (do_sample=False, num_beams=1, temperature=0)
• vllm == 0.3.0 does not support Gemma and Apollo. vllm==0.3.2 does.
• Under multi-gpu setting, when you run eval_bench.py with --use_vllm, you might face the error RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method (for example when using vllm==0.6.1.post2.) If so, please add the environmental variable with a line of code os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn".

Environment

Acknowledgement / 謝辞

This work was supported by AIST KAKUSEI project (FY2023).
本研究は、国立研究開発法人産業技術総合研究所事業の令和5年度覚醒プロジェクトの助成を受けたものです。

MedMCQA and MedQA were provided at JMedBench by Junfeng Jiang.

How to cite

Please cite our paper if you use this code!

@article{sukeda2024development,
  title={{Development and bilingual evaluation of Japanese medical large language model within reasonably low computational resources}},
  author={Sukeda, Issey},
  journal={arXiv preprint arXiv:2409.11783},
  year={2024},
}