Training Language Models to Identify Urgent Messages in Real-Time
summary
Fine-tunes Gemma-2 2B instruction model on a synthetic dataset of 13,000 synthetic messages (12k train, 1k validation) to classify urgent vs non‑urgent maternal health queries for MomConnect, producing structured JSON with a probability and a reason-first explanation. Training uses LoRA, 4‑bit quantization experiments, and yields accuracy comparable to GPT‑3.5‑turbo while matching or beating some baselines; total compute is about 36–48 GPU hours, data curation cost ≈ $50, and the deployed urgency filter forwards truly urgent messages to human help desks.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:06Speaker background
- 2:33MomConnect challenges
- 3:55Urgent message solution
- 4:58Data challenges
- 6:24Synthetic personas
- 8:33Synthetic message generation
- 11:02Language model selection
- 13:08Model input/output
- 14:32Hyperparameter considerations for training
- 16:22Model performance
- 18:09Key takeaways
- 20:36Model limitations
- 22:20Scaling challenges
- 23:09Q&A
related talks
09mUsing and fine-tuning Gemma 3
10mDemo: Using Gemma with the Hugging Face ecosystem
14mGemma on mobile and web. Best and worst practices
12mGemma: The responsible way to build
01h 21mTLMs: Tiny LLMs and Agents on Edge Devices with LiteRT-LM
44mBuilding amazing Front-End Project with Gemma: A State-of-the-Art Open Model
33mLessons Learned Applying Large Language Models in Healthcare
10mGemma 3 architecture and design