ELAINE-medllm - Build with Llama3-8B
ELAINE (EngLish-jApanese-chINesE)-medLLM is a trilingual (English, Japanese, Chinese) large language mol adapted for the bio-medical domain based on Llama-3-8B. The training dataset was carefully curated in terms of volume and diversity to adapt to the biomedical domain and endow trilingual capability while preserving the knowledge and abilities of the base model. The training follows 2-stage paths: continued pre-training and supervised fine-tuning (SFT). ELAINE-medLLM exhibits superior trilingual capabilities compared to existing bilingual or multilingual medical LLMs without severely sacrificing the base model's capability.
Model Details
- Model type: Please refer to Llama 3 Github for details on the model architecture.
- Language(s): English, Japanese, Chinese
- Library: DeepSpeed
- Tokenizer: Please refer to Llama 3 blog for details on the tokenizer.
Model Performance
Evaluation Benchmarks
The evaluation behchmark dataset and evaluation code can be obtained from this Github site. The details of the bechmark are as follows.
English evaluation benchmarks
Japanese evaluation benchmarks
- IgakuQA
- We concatenate the original exam data from 2018 to 2022 into a single JSON file.
- JJSIMQA
- DenQA
- It contains the exam problems from the Japan National Dentistry Examination and their answers in the past two years (from 2023 through 2024) extracted from the official website of the Ministry of Health, Labor and Welfare in Japan (https://www.mhlw.go.jp/stf/english/index.html).
Chinese evaluation benchmarks
Training Datasets
Continued pre-training
For continued pretraining, we collected English, Japanese, and Chinese text in the bio-medical domain. The domain text collected is classified into six categories: 1) scientific papers, 2) medical guidelines, 3) web text related to biomedical, 4) textbook of biomedical, 5) PubMed abstracts, and 6) PubMed Central (PMC) archives. For the Japanese PubMed abstract, we used the original English PubMed abstract translated in Japanese. We used only open-licensed text except for the Japanese biomedical papers from J-STAGE.
Instruction supervised fine-tuning
We collected various conversational QA datasets in the bio-medical domain from different data sources. For English, we used Medical Meadow in MedAlpca, HealthCareMagic, and iClilic dataset used in ChatDoctor. We adapted the augmented QA dataset from HuatuoGPT-2 for Chinese and English. For Japanese, we used existing alpaca datasets in the general domain translated in Japanese.
Results
English benchmark
model_name | MMLU | MedMCQA | MedQA | MedQA-4op | PubMedQA | Avg |
---|---|---|---|---|---|---|
google_gemma-7b-it | 50.55 | 41.07 | 33.12 | 39.67 | 67.07 | 46.30 |
meta-llama_Llama-2-7b-chat-hf | 48.71 | 35.97 | 30.99 | 38.09 | 63.64 | 43.48 |
meta-llama_Meta-Llama-3-8B-Instruct | 72.79 | 60.89 | 57.65 | 61.28 | 78.99 | 66.32 |
tokyotech-llm_Llama-3-Swallow-8B-Instruct-v0.1 | 66.88 | 53.85 | 47.95 | 56.07 | 64.65 | 57.88 |
medalpaca_medalpaca-7b | 51.48 | 36.02 | 31.15 | 39.35 | 55.15 | 42.63 |
epfl-llm_meditron-7b | 47.32 | 34.35 | 29.18 | 32.26 | 39.19 | 36.46 |
aaditya_Llama3-OpenBioLLM-8B | 73.43 | 55.03 | 50.00 | 56.78 | 65.86 | 60.22 |
FreedomIntelligence_Apollo-7B | 68.17 | 53.85 | 45.98 | 53.86 | 75.35 | 59.44 |
Llama3-ELAINE-medLLM-instruct-8B | 72.69 | 55.07 | 55.76 | 61.36 | 75.35 | 64.05 |
Japanese benchmark
model_name | DenQA | IgakuQA | JJSIMQA | Avg |
---|---|---|---|---|
google_gemma-7b-it | 13.71 | 25.51 | 12.09 | 17.10 |
meta-llama_Llama-2-7b-chat-hf | 12.03 | 20.80 | 10.55 | 14.46 |
meta-llama_Meta-Llama-3-8B-Instruct | 19.72 | 40.45 | 25.93 | 28.70 |
tokyotech-llm_Llama-3-Swallow-8B-Instruct-v0.1 | 23.78 | 44.01 | 26.81 | 31.53 |
medalpaca_medalpaca-7b | 10.91 | 17.74 | 10.77 | 13.14 |
epfl-llm_meditron-7b | 9.79 | 18.20 | 8.35 | 12.11 |
aaditya_Llama3-OpenBioLLM-8B | 18.18 | 33.03 | 21.98 | 24.40 |
FreedomIntelligence_Apollo-7B | 17.90 | 32.28 | 20.66 | 23.61 |
Llama3-ELAINE-medLLM-instruct-8B | 22.24 | 43.36 | 24.40 | 30.00 |
Chinese benchmark
model_name | CMExam | MedQA | MedQA-4op | Avg |
---|---|---|---|---|
google_gemma-7b-it | 30.90 | 29.03 | 34.96 | 31.63 |
meta-llama_Llama-2-7b-chat-hf | 25.43 | 25.37 | 32.30 | 27.70 |
meta-llama_Meta-Llama-3-8B-Instruct | 52.01 | 62.99 | 68.40 | 61.13 |
tokyotech-llm_Llama-3-Swallow-8B-Instruct-v0.1 | 41.11 | 45.05 | 51.27 | 45.81 |
medalpaca_medalpaca-7b | 23.58 | 24.99 | 30.11 | 26.23 |
epfl-llm_meditron-7b | 23.85 | 25.46 | 29.82 | 26.38 |
aaditya_Llama3-OpenBioLLM-8B | 39.07 | 42.59 | 48.73 | 43.46 |
FreedomIntelligence_Apollo-7B | 49.99 | 58.29 | 62.99 | 57.09 |
Llama3-ELAINE-medLLM-instruct-8B | 48.85 | 55.80 | 61.59 | 55.41 |
Risks and Limitations
The models released here are still in the early stages of our research and development and have not been tuned to ensure outputs align with human intent and safety considerations.
Acknowledgements
We thank Meta Research for releasing Llama 3 under a generous open license.
Authors
- Ken Yano
- Zheheng Luo
- Jimin Huang
- Qianqian Xie
- Masaki Asada
- Chenhan Yuan
- Kailai Yang
- Makoto Miwa
- Sophia Ananiadou
- Jun'ichi Tsujii
Contact
- Ken Yano [[email protected]]
How to cite
If you find our work helpful, please feel free to cite these papers.
@article{published_papers/48577159,
title = {ELAINE-medLLM: Lightweight English Japanese Chinese Trilingual Large Language Model for Bio-medical Domain (To appear)},
author = {Ken Yano and Zheheng Luo and Jimin Huang and Qianqian Xie and Masaki Asada and Chenhan Yuan and Kailai Yang and Makoto Miwa and Sophia Ananiadou and Jun'ichi Tsujii},
journal = {The 31st International Conference on Computational Linguistics (COLING 2025)},
month = {1},
year = {2025}
}
- Downloads last month
- 93
Model tree for kenyano/Llama3-ELAINE-medLLM-instruct-8B
Base model
meta-llama/Meta-Llama-3-8B