MedSwallow-70b

⚠️⚠️⚠️ Only for research purpose. Do not use it for medical purpose. ⚠️⚠️⚠️

MedSwallow-70B🏥

東工大Swallowをベースモデルとし, 医療Q&AデータセットでInstruction Tuningを施した医療ドメインの日本語LLMです.

チューニングには独自で用意した米国医師国家試験(USMLE)を和訳したQ&Aデータセットを用いました.

MedSwallow is a Japanese medical LLM for medical question-answering.

MedSwallow is based on Swallow-70B and has passed instruction tuning with USMLE dataset translated in Japanese by our own.

Training procedure

The following bitsandbytes quantization config was used during training:

• quant_method: bitsandbytes
• load_in_8bit: False
• load_in_4bit: True
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: nf4
• bnb_4bit_use_double_quant: True
• bnb_4bit_compute_dtype: bfloat16

Framework versions

• PEFT 0.4.0

License

ライセンスは非商用ライセンスです.

Non-commercial.

Usage

model_name = "tokyotech-llm/Swallow-70b-instruct-hf"
peft_model= "AIgroup-CVM-utokyohospital/MedSwallow-70b"

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

bnb_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=torch.float16,
        )

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    load_in_8bit=False,
    torch_dtype=torch.float16,
    device_map=device,
        
model = PeftModel.from_pretrained(
    model, 
    peft_model, 
    torch_dtype=torch.float16,
    device_map=device, 
)

Benchmark

See also Japanese Medical Language Model Evaluation Harness.

• IgakuQA (in English):
• IgakuQA (in Japanese):
• MedQA (in English) :
• MedQA (in Japanese) :

How to cite

@misc{sukeda202470bparameterlargelanguagemodels,
      title={70B-parameter large language models in Japanese medical question-answering}, 
      author={Issey Sukeda and Risa Kishikawa and Satoshi Kodera},
      year={2024},
      eprint={2406.14882},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2406.14882}, 
}