Coursera Spring Sale
40% Off Coursera Plus Annual!
Grab it
Explore the fundamental flaws in current fine-tuning practices through this 25-minute conference talk that challenges conventional approaches to model optimization. Learn why simply adding more data with the same pretraining loss constitutes extended pretraining rather than true fine-tuning, and discover how this misunderstanding leads to brittle, inefficient models in production environments. Examine the critical differences between superficial fine-tuning and genuine model adaptation that involves modifying loss functions, adjusting output heads, and optimizing for real-world constraints including confidence calibration, consistency, and latency requirements. Gain insights into strategic decision-making around when to implement fine-tuning versus alternative approaches, and understand the substantial advantages of proper fine-tuning techniques that extend beyond mere stylistic adjustments to encompass output constraints, confidence scoring, and dramatic latency improvements. Discover practical strategies for tailoring models to meet specific production needs while avoiding common pitfalls that compromise model performance and reliability in deployed systems.
