|
--- |
|
license: cc |
|
language: |
|
- en |
|
library_name: transformers |
|
pipeline_tag: text-generation |
|
tags: |
|
- medical |
|
--- |
|
# MedAlpaca 7b |
|
|
|
|
|
## Table of Contents |
|
|
|
[Model Description](#model-description) |
|
- [Architecture](#architecture) |
|
- [Training Data](#trainig-data) |
|
[Model Usage](#model-usage) |
|
[Limitations](#limitations) |
|
|
|
## Model Description |
|
### Architecture |
|
`medalpaca-7b` is a large language model specifically fine-tuned for medical domain tasks. |
|
It is based on LLaMA (Large Language Model Meta AI) and contains 7 billion parameters. |
|
The primary goal of this model is to improve question-answering and medical dialogue tasks. |
|
Architecture |
|
|
|
|
|
### 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](https://www.wikidoc.org/index.php/Main_Page). |
|
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](https://arxiv.org/abs/2303.14070) |
|
consisting of 200,000 question-answer pairs, available at https://github.com/Kent0n-Li/ChatDoctor. |
|
|
|
| Source | n items | |
|
|------------------------------|--------| |
|
| ChatDoc large | 200000 | |
|
| wikidoc | 67704 | |
|
| Stackexchange academia | 40865 | |
|
| Anki flashcards | 33955 | |
|
| Stackexchange biology | 27887 | |
|
| Stackexchange fitness | 9833 | |
|
| Stackexchange health | 7721 | |
|
| Wikidoc patient information | 5942 | |
|
| Stackexchange bioinformatics | 5407 | |
|
|
|
## Model Usage |
|
To evaluate the performance of the model on a specific dataset, you can use the Hugging Face Transformers library's built-in evaluation scripts. Please refer to the evaluation guide for more information. |
|
Inference |
|
|
|
You can use the model for inference tasks like question-answering and medical dialogues using the Hugging Face Transformers library. Here's an example of how to use the model for a question-answering task: |
|
|
|
```python |
|
|
|
from transformers import pipeline |
|
|
|
pl = pipeline("text-generation", model="medalpaca/medalpaca-7b", tokenizer="medalpaca/medalpaca-7b") |
|
question = "What are the symptoms of diabetes?" |
|
context = "Diabetes is a metabolic disease that causes high blood sugar. The symptoms include increased thirst, frequent urination, and unexplained weight loss." |
|
answer = pl(f"Context: {context}\n\nQuestion: {question}\n\nAnswer: ") |
|
print(answer) |
|
``` |
|
|
|
## 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. |
|
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard) |
|
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_medalpaca__medalpaca-7b) |
|
|
|
| Metric | Value | |
|
|-----------------------|---------------------------| |
|
| Avg. | 44.98 | |
|
| ARC (25-shot) | 54.1 | |
|
| HellaSwag (10-shot) | 80.42 | |
|
| MMLU (5-shot) | 41.47 | |
|
| TruthfulQA (0-shot) | 40.46 | |
|
| Winogrande (5-shot) | 71.19 | |
|
| GSM8K (5-shot) | 3.03 | |
|
| DROP (3-shot) | 24.21 | |
|
|