medAlpaca: Finetuned Large Language Models for Medical Question Answering

Project Overview

MedAlpaca expands upon both Stanford Alpaca and
AlpacaLoRA to offer an advanced suite of large language
models specifically fine-tuned for medical question-answering and dialogue applications.
Our primary objective is to deliver an array of open-source language models, paving the way for
seamless development of medical chatbot solutions.

These models have been trained using a variety of medical texts, encompassing resources such as
medical flashcards, wikis, and dialogue datasets. For more details on the data utilized, please consult the data section.

Getting Started

Create a new virtual environment, e.g. with conda

conda create -n medalpaca python>=3.9

Install the required packages:

pip install -r requirements.txt

Training of medAlpaca

Memory Requirements

We have benchmarked the needed GPU memory as well as the approximate duration per epoch
for finetuning LLaMA 7b on the Medical Meadow small dataset (~6000 Q/A pairs) on a single GPU:

Train medAlpaca based on LLaMA

If you have access to the LLaMA or Alpaca
weights you can finetune the model with the following command.
Just replace <PATH_TO_LLAMA_WEIGHTS> with the folder containing you LLaMA or Alpaca weights.

python medalpaca/train.py \
    --model PATH_TO_LLAMA_WEIGHTS \
    --data_path medical_meadow_small.json \
    --output_dir 'output' \
    --train_in_8bit True \  
    --use_lora True \
    --bf16 True \
    --tf32 False \
    --fp16 False \
    --global_batch_size 128 \
    --per_device_batch_size 8 \

Per default the script performs mixed precision training.
You can toggle 8bit training with the train_in_8bit flag.
While 8 bit training currently only works with use_lora True, however you can use
LoRA without 8 bit training.
It is also able to train other models such as facebook/opt-6.7 with the above script.

Data

To ensure your cherished llamas and alpacas are well-fed and thriving,
we have diligently gathered high-quality biomedical open-source datasets
and transformed them into instruction tuning formats.
We have dubbed this endeavor Medical Meadow.
Medical Meadow currently encompasses roughly 1.5 million data points across a diverse range of tasks,
including openly curated medical data transformed into Q/A pairs with OpenAI's gpt-3.5-turbo
and a collection of established NLP tasks in the medical domain.
Please note, that not all data is of the same quantitiy and quality and you may need tp subsample
the data for training your own model.
We will persistently update and refine the dataset, and we welcome everyone to contribute more 'grass' to Medical Meadow!

Data Overview

Data description

please refer to DATA_DESCRIPTION.md

Benchmarks

We are benchmarking all models on the USMLE self assessment, which is available at this link.
Note, that we removed all questions with images, as our models are not multimodal.

We are continuously working on improving the training as well as our evaluation prompts.
Expect this table to change quite a bit.

Access the models

Visit the zoo and have a look at our alpacas here: https://huggingface.co/medalpaca

It should be obvious, but the models provided on this platform are shared for research purposes
only and should not be used in any healthcare applications or settings.
While we are excited to showcase our experimental models, please be aware that they have not undergone
extensive testing or validation, and their reliability cannot be guaranteed.
We kindly ask you to exercise caution when using these models,
and we appreciate your understanding as we continue to explore and develop this innovative technology.

Paper

@article{han2023medalpaca,
  title={MedAlpaca--An Open-Source Collection of Medical Conversational AI Models and Training Data},
  author={Han, Tianyu and Adams, Lisa C and Papaioannou, Jens-Michalis and Grundmann, Paul and Oberhauser, Tom and L{\"o}ser, Alexander and Truhn, Daniel and Bressem, Keno K},
  journal={arXiv preprint arXiv:2304.08247},
  year={2023}
}