Introducing the PULSE Project
PULSE is an innovative project that stands out in the field of medical AI with its specialized focus on Chinese medical language models. It represents a significant advancement in natural language processing (NLP) for healthcare applications and is designed to address various tasks within the medical domain.
Main Features of the PULSE Model
PULSE is a Chinese medical large language model that has undergone extensive training. The model has been fine-tuned using approximately 4 million instructions from both medical and general Chinese domains. This comprehensive training allows PULSE to excel at numerous medical NLP tasks, which include:
- Health education
- Passing medical exams
- Evaluating and interpreting medical reports
- Structuring medical records
- Simulating diagnostic and treatment processes
Model Availability
PULSE is available in different model sizes, which cater to various needs:
- PULSE-7b is built on the bloomz-7b1-mt architecture.
- PULSE-20b utilizes the InternLM-20B base for fine-tuning.
Moreover, a quantized version of the model will soon be available, catering to those needing larger parameter models. The team encourages partnerships for further expansion.
Limitations
Despite its powerful capabilities, PULSE is primarily for research use in the medical field. Its results should not replace professional medical advice, as the model's information may not always be complete or fully accurate. Users are advised to consult professional healthcare providers for diagnosis and treatment.
Elo Evaluation
The performance of PULSE is evaluated using the Elo Rating tournament method, which helps balance the cost of model assessment. PULSE has shown competitive results in several benchmark datasets, including MedQA and PromptCBLUE, among others.
Additional Resources
For further model fine-tuning, PULSE integrates capabilities from the LLaMA-Factory project, and guidance is available for quantizing the model through LMDeploy.
Inference Requirements
To deploy PULSE locally for inference with a batch size of 1, users need the following hardware capacities:
- 7B Model: 14GB for FP16 or 6GB for INT4
- 20B Model: 40GB for FP16 or 12GB for INT4
Download and Installation
To begin using PULSE, follow these steps:
-
Clone the repository:
git clone https://github.com/openmedlab/PULSE cd PULSE
-
Set up a conda environment and install dependencies:
conda env create -f llm.yml conda activate llm
Usage Examples
PULSE can be used to perform various tasks, including:
- Health education demonstrations
- Medical examination question answering
- Interpreting medical reports
- Structuring medical records
- Simulating consultations
It also effectively handles non-medical inquiries in a safe manner.
Related Projects
The PULSE project is connected with several initiatives, exploring various applications such as X-ray analysis with language models and COVID-19 specific adaptations. Other related projects enhance medical terminology normalization and provide libraries for healthcare chatbot development.
Acknowledgments
PULSE was developed with contributions from notable organizations, including the Shanghai Artificial Intelligence Laboratory and universities specializing in natural language processing and data mining.
License
The project's code is available under the Apache 2.0 license, and the model weights are shared under the GNU AGPL 3.0 license.
PULSE represents a leap forward in the use of AI models for healthcare, offering a diverse range of applications that promise to advance medical understanding and delivery.