Statistical Insight from PubMedQA

Inspiration

Our teams inspiration was sought through innovative thinking and collaborative problem solving. We were looking to take a creative approach towards a technical data problem.

What it does

Runs a statistical analysis on a PubMedQA dataset to identify where abstracts may lead an LLM or any reader into a direction that is not in line with an expert response. To go further, we developed a tool that is utilized by an LLM to generate variable maps to generate a grounded response to increase accuracy of a question.

How we built it

We used Python for the statistical analysis and google adk for the development and deployment of the tool.

Challenges we ran into

One challenge our team faced was coding errors when developing our tool, and visualization errors when creating statistical plots.

Accomplishments that we're proud of

We are proud to have completed our first BNFOthon and figured out how to create creative problem solving when developing a technical problems. We were able to generate grounded responses and grounded calculations and achieve traceability in generations.

What we learned

We learned how to run statistical analysis on natural language and think critically about how hedging language influence LLM accuracy alongside viewer response.