AIgroup-CVM-utokyohospital/llama2-jmedlora-30000

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

This model is an instruction-tuned model of Llama2-70B with our own medical Q&A dataset.

Method

QLoRA

Parameters

• batch_size = 512
• max_steps = 30000 (around 6.89 epochs)
• source_max_len = 512
• target_max_len = 512

Training time

1617017 seconds on NVIDIA A100 x 4 (not fully used)

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

How to cite

本データを利用する場合は以下の文献の引用をご検討ください.

@article{sukeda2023jmedlora,
  title={{JMedLoRA: Medical Domain Adaptation on Japanese Large Language Models using Instruction-tuning}},
  author={Sukeda, Issey and Suzuki, Masahiro and Sakaji, Hiroki and Kodera, Satoshi},
  journal={arXiv preprint arXiv:2310.10083},
  year={2023}
}