Constructing Synthetic Datasets Using LLMs

In this post you will read about LLMs for synthetic data to reduce bias.

In our recent LLM workshop Matt McInnis showcased the use of large language models and synthetic data sets to address the digital divide and create inclusive educational tools. He highlighted the challenges faced during implementation and the improvements made through a multi-step pipeline. The benefits of using large language models, such as explainability and increased productivity, were emphasized. Matt's company, Typist, aims to bridge the gap between technology access and education, ultimately making a difference in the lives of those they serve.

Topics:
-------
⃝ Addressing the Digital Divide

* Typist's mission is to reduce the digital divide by providing easy-to-use technology solutions
* Large language models are important in creating a medical office educational simulator
* The simulator requires the construction of complex synthetic data sets


⃝ Ensuring Representation and Diversity

* Generated names in the synthetic data sets were not representative of the student population or the community they serve
* Ontario census data was used as a benchmark for ethnic and cultural diversity
* OpenAI's GPT 3.5 turbo was used to predict the ethnic or cultural origins associated with each name


⃝ Implementation Challenges and Improvements

* Initial approach using AI labeler and GPT 3.5 turbo had limitations in determining name origins accurately
* Response times were longer than desired
* Multi-step pipeline incorporating generative knowledge prompting and chain of thought prompting techniques improved accuracy


⃝ Benefits of Large Language Models

* Explainability provided by LLMs allows for backing up generated data with reasoning
* LLMs are cost-effective and easy to implement, improving development velocity and productivity
* LLMs have the potential to improve the efficiency of medical clinics and reduce negative patient outcomes

---