12-03, 16:30–17:00 (UTC), LLM Track
Providing timely maternal healthcare in developing countries is a critical challenge. This talk demonstrates how data-driven solutions can bridge healthcare gaps and improve access to vital healthcare information for pregnant women, with user privacy in mind. To do so, we fine-tuned the Gemma-2 2 billion parameter instruction model on a synthetic dataset in order to detect whether user messages pertain to urgent or non-urgent maternal healthcare issues. By quickly identifying and prioritizing user inquiries, the model can aid help desks by ensuring urgent messages are promptly forwarded to the appropriate healthcare professionals for immediate intervention.
This talk will walk the audience through the following areas:
- The status of maternal healthcare in developing countries and why a custom, privacy-preserving solution is often needed. (~2 - 3 min)
- How to think about (and carry out) data curation using large language models in order to train smaller, custom language models that can be deployed to edge devices. (10 min)
- How to select an appropriate small language model for fine-tuning on the curated dataset. (5 min)
- The details involved in training the small language model and resources required. (5 min)
- Our main results, comparing accuracy and AUC against OpenAI models such as GPT-3.5-turbo, GPT-4o-mini, etc. (~2 - 3 min)
- Important lessons learned from the data curation process and how our custom model can be further improved. (5 min)
At the end of the talk, the audience will better understand how to think about data curation using large language models for training smaller language models and some of the common pitfalls in the overall process.
The audience is expected to have general knowledge regarding large language models. Knowledge on model training is helpful, but not required.
No previous knowledge expected
Tony has a broad background in the automotive and tech industries , with a focus on deep learning, particularly within the natural language processing domain. He has worked extensively on training language models, orchestrating GenAI ecosystems, and developing tools for LLM evaluation, decision-making, and hallucination mitigation. His experience ranges from hardware testing with LiDARs on autonomous vehicles to object recognition using computer vision and generative AI applications.