medguanaco-lora-33b-8bit

Medguanaco LoRA 33b 8bit

Table of Contents

Model Description

• Architecture
• Training Data
  Model Usage
  Limitations

Model Description

Architecture

nmitchko/medguanaco-lora-33b-8bit is a large language model LoRa specifically fine-tuned for medical domain tasks. It is based on the Guanaco LORA of LLaMA weighing in at 33B parameters. The primary goal of this model is to improve question-answering and medical dialogue tasks. It was trained using LoRA and reduced to 8bit, to reduce memory footprint.

Steps to load this model:

1. Load Guanaco-33b-merged https://huggingface.co/timdettmers/guanaco-33b-merged in 8-bit
   • I recommend using text-generation-ui to test it out: https://github.com/oobabooga/text-generation-webui/tree/main
2. Apply this LoRA to the model, this was trainied in 8-bit mode and results may vary in higher dimensions.

# Some llama or alpaca model 65b
base_model = "timdettmers/guanaco-33b-merged"
model = LlamaForCausalLM.from_pretrained(
    base_model,    
    load_in_8bit=load_8bit,
    torch_dtype=torch.float16
)
# Load the LORA on top
lora_weights = "nmitchko/medguanaco-lora-33b-8bit"
model = PeftModel.from_pretrained(
    model,
    lora_weights,
    torch_dtype=torch.float16
)

---

The following README is taken from the source page medalpaca

Training Data

The training data for this project was sourced from various resources. Firstly, we used Anki flashcards to automatically generate questions, from the front of the cards and anwers from the back of the card. Secondly, we generated medical question-answer pairs from Wikidoc. We extracted paragraphs with relevant headings, and used Chat-GPT 3.5 to generate questions from the headings and using the corresponding paragraphs as answers. This dataset is still under development and we believe that approximately 70% of these question answer pairs are factual correct. Thirdly, we used StackExchange to extract question-answer pairs, taking the top-rated question from five categories: Academia, Bioinformatics, Biology, Fitness, and Health. Additionally, we used a dataset from ChatDoctor consisting of 200,000 question-answer pairs, available at https://github.com/Kent0n-Li/ChatDoctor.

Limitations

The model may not perform effectively outside the scope of the medical domain. The training data primarily targets the knowledge level of medical students, which may result in limitations when addressing the needs of board-certified physicians. The model has not been tested in real-world applications, so its efficacy and accuracy are currently unknown. It should never be used as a substitute for a doctor's opinion and must be treated as a research tool only.