About SyDeTre

Hi, I'm Toni, and this is my project called SyDeTre, a medical model finetuned for symptom detection and treatment.

Project Motivation

I chose this task because pharmaceutical drugs and various treatments are often missing in medicinal LLM datasets. To address this gap, I started by gathering a comprehensive multi-point database containing thousands of illnesses and their corresponding medications.

Data Collection and Processing

I began by querying questions to these data points using Mistral 8x22b and incorporated data from Medquad. This effort resulted in an MIT Licensed medical dataset consisting of 72,000 Q&A pairs.

Finetuning Approach

For finetuning the model, I experimented with different hyperparameters, finding that a higher number of training steps and a smaller learning rate were beneficial. By iterating the process based on wandb results, I achieved positive outcomes.

Deployment

The final model was deployed with a Next.js interface, allowing users to chat with SyDeTre and utilize its medical insights.

Future Plans

I am pleased with the results of the dataset and finetuning efforts. I plan to continue improving the model and will test SyDeTre on medical LLM benchmarks in the future.

Conclusion

This project provided valuable insights and a unified approach to finetuning with the Mistral API. I am grateful for the opportunity to share my work.

Built With

Share this project:

Updates