Fine-Tuning BERT for the Unstructured Data You Actually Have

Resources: code: https://github.com/unionai/workshops/tree/main/tutorials/bert-fine-tuning-emotion Flyte: https://github.com/flyteorg/flyte Slack: https://slack.flyte.org/ Connect with Sage Elliott: https://www.linkedin.com/in/sageelliott/ ​Most fine-tuning attention goes to generative LLMs, but a large share of production NLP still runs on BERT-family encoders. They are small, fast, and cheap to serve, and on the tasks where most real data lives (classifying support tickets, extracting fields from documents, routing emails, semantic search) a fine-tuned BERT often matches or beats a prompted frontier model at a fraction of the cost and latency. What we'll cover - ​Where encoder models like BERT fit, and why they still win on classification, extraction, and embedding tasks ​- Fine-tuning an open-weight BERT model with Hugging Face Transformers ​- Orchestrating with Flyte 2: cached data prep, GPU-aware training, reproducible runs at any scale ​Deploying behind a UI, with a path to low-latency, scaled inference