-
-
-
5th answer from fine-tuned dolly
-
5 customized questions and their answers from original dolly
-
Answers 1-4 part-1 (from fine-tuned dolly) for comparison
-
Answers 1-4 part-2 (from fine-tuned dolly) for comparison
Inspiration
We were inspired by the growing need for accessible and reliable healthcare information. We observed how people often faced challenges in finding accurate medical advice promptly, leading to confusion and unnecessary delays in seeking appropriate care. This inspired us to develop a solution that could bridge the gap and empower individuals with easy access to personalized healthcare guidance.
What it does
MediTalk is a Q&A chatbot in the domain of healthcare. It is created to solve the queries of distressed people wanting instant replies to their problems relating to their medication or symptoms. This chatbot is very easy to use and is trained in simple language to facilitate every person. Throughout the development process, We immersed ourselves in extensive research on medical resources, AI algorithms, natural language processing, and large language models. We learned about the complexities of medical diagnoses and treatments, ensuring that MediTalk could provide accurate information and helpful recommendations. Collaborating with healthcare professionals and experts in the field, we gained valuable insights into the nuances of various medical conditions, symptom analysis, and evidence-based practices.
How we built it
To build MediTalk, we utilized advanced AI technologies and programming languages. We first analyzed and compared different medical datasets online and chose the most relevant one amongst a huge collection of datasets. Then we chose the free and open source large language model to be fine-tuned- databricks/dolly-v2-3b (causal language model) for the required task as we did not find any of its implementations on a similar dataset to date. We chose the 3 billion parameter due to a lack of computational resources. Finally, we analyzed and compared the performance of the fine-tuned model with the original dolly and found it to be much more precise and extremely faster (about 10 times) on medical questioning.
Challenges we ran into
One of the major challenges we faced was a lack of computational resources, we utilized a single Tesla T4 GPU with 16GB GPU RAM and 12 GB CPU RAM. The additional runtime limits of google colab worsened the problem. Although, we overcame these issues by implementing relevant strategies and striving hard to accomplish the best possible results in the given timeframe.
Accomplishments that we're proud of
We created the foundation of a novel medical question-answering chatbot called MediTalk, utilizing the open-source LLM dolly. This chatbot delivers accurate responses within a matter of seconds (10-15 seconds) and has the potential to revolutionize the field of medical diagnosis and healthcare when trained on high-performance hardware such as advanced GPUs and RAM.
What we learned
This journey has proved to be very informative and has helped all of us to up our game and have a more clear view of our goals. This hackathon has provided us with a deep insight into the architecture as well as the model building, training, and tuning of the model chosen, which is Dolly, in our case. The most time-consuming part was the dataset searching which took quite some time. After that, The struggle to integrate and finetune the model was the real task at hand. We did not give up and pushed ourselves further to test our limits to get the appropriate output. From hours of struggling through crashing RAMs to finally fine-tuning, saving, and implementing our fine-tuned version of Dolly, it was indeed a mesmerizing journey.
What's next for MediTalk
Building MediTalk has been a remarkable journey, and we are proud to witness the positive impact it might have on people's lives. With each interaction and feedback received, we will continue to refine and improve MediTalk, ensuring that it achieves ground-breaking results in medical question answering, helping in spreading awareness, and providing easy preventive diagnosis for all. We believe that MediTalk is highly scalable as the field of computational biology arises.
Log in or sign up for Devpost to join the conversation.