MMedLM

The official model weights for "Towards Building Multilingual Language Model for Medicine".

Introduction

This repo contains MMedLM 2-1.8B , a multilingual medical foundation model with 1.8 billion parameters. MMedLM 2-1.8B builds upon the foundation of InternLM 2-1.8B and has been further pretrained on MMedC, a comprehensive multilingual medical corpus. This further pretraining enhances the model's medical-domain knowledge. With an auto-regressive continues training on MMedC, MMedLM 2-1.8B can exceed the performance of most 7B models, including InternLM and LLaMA 2.

The model underwent further pretraining on MMedC with the following hyperparameters:

Iterations: 15000
Global batch size: 512
Cutoff length: 2048
Learning rate: 2e-5

The model can be loaded as follows:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("Henrychur/MMedLM2-1.8B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("Henrychur/MMedLM2-1.8B", torch_dtype=torch.float16, trust_remote_code=True)

Note that this is a foundation model that has not undergone instruction fine-tuning.

News

[2023.3.1] We release MMedLM 2-1.8B, a 1.8B light-weight model based on InternLM 2-1.8B. With an auto-regressive continues training on MMedC, MMedLM 2-1.8B can exceed the performance of most 7B models, including InternLM and LLaMA 2.

[2024.2.21] Our pre-print paper is released ArXiv. Dive into our findings here.

[2024.2.20] We release MMedLM and MMedLM 2. With an auto-regressive continues training on MMedC, these models achieves superior performance compared to all other open-source models, even rivaling GPT-4 on MMedBench.

[2023.2.20] We release MMedC, a multilingual medical corpus containing 25.5B tokens.

[2023.2.20] We release MMedBench, a new multilingual medical multi-choice question-answering benchmark with rationale. Check out the leaderboard here.

Evaluation on MMedBench

The further pretrained MMedLM 2 showcast it's great performance in medical domain across different language.

Method	Size	Year	MMedC	MMedBench	English	Chinese	Japanese	French	Russian	Spanish	Avg.
GPT-3.5	-	2022.12	✗	✗	56.88	52.29	34.63	32.48	66.36	66.06	51.47
GPT-4	-	2023.3	✗	✗	78.00	75.07	72.91	56.59	83.62	85.67	74.27
Gemini-1.0 pro	-	2024.1	✗	✗	53.73	60.19	44.22	29.90	73.44	69.69	55.20
BLOOMZ	7B	2023.5	✗	trainset	43.28	58.06	32.66	26.37	62.89	47.34	45.10
InternLM	7B	2023.7	✗	trainset	44.07	64.62	37.19	24.92	58.20	44.97	45.67
Llama 2	7B	2023.7	✗	trainset	43.36	50.29	25.13	20.90	66.80	47.10	42.26
MedAlpaca	7B	2023.3	✗	trainset	46.74	44.80	29.64	21.06	59.38	45.00	41.11
ChatDoctor	7B	2023.4	✗	trainset	43.52	43.26	25.63	18.81	62.50	43.44	39.53
PMC-LLaMA	7B	2023.4	✗	trainset	47.53	42.44	24.12	20.74	62.11	43.29	40.04
Mistral	7B	2023.10	✗	trainset	61.74	71.10	44.72	48.71	74.22	63.86	60.73
InternLM 2	1.8B	2024.2	✗	trainset	38.49	64.1	32.16	18.01	53.91	36.83	40.58
InternLM 2	7B	2024.2	✗	trainset	57.27	77.55	47.74	41.00	68.36	59.59	58.59
MMedLM (Ours)	7B	-	✓	trainset	49.88	70.49	46.23	36.66	72.27	54.52	55.01
MMedLM 2(Ours)	7B	-	✓	trainset	61.74	80.01	61.81	52.09	80.47	67.65	67.30
MMedLM 2(Ours)	1.8B	-	✓	trainset	45.40	66.78	42.21	25.56	69.14	43.40	48.75

GPT and Gemini is evluated under zero-shot setting through API
Open-source models first undergo training on the trainset of MMedBench before evaluate.

Contact

If you have any question, please feel free to contact [email protected].

Citation

@misc{qiu2024building,
      title={Towards Building Multilingual Language Model for Medicine}, 
      author={Pengcheng Qiu and Chaoyi Wu and Xiaoman Zhang and Weixiong Lin and Haicheng Wang and Ya Zhang and Yanfeng Wang and Weidi Xie},
      year={2024},
      eprint={2402.13963},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Henrychur
/

MMedLM2-1_8B