Episode Details
Back to Episodes
When Fine-Tuning Over-Generalizes Instead of Forgetting
Episode 6543
Published 2 days, 7 hours ago
Description
Fine-tuning a speech model on IP addresses can make ordinary phrases like "ten one" transcribe as "10.1." Rather than catastrophic forgetting, this quieter failure is over-generalization, where a model applies a new surface pattern everywhere. The discussion unpacks why labels act as a teaching surface, how to catch formatting bleed using contrast sets and B-WER/U-WER splits, and what replay-share sweeps from Ibrahim and colleagues actually show about preventing forgetting.
Episode 6543 — myweirdprompts.com/n/6543 (permanent short link: myweirdprompts.com/661228)
Listen on Android: MWP FM Player — myweirdprompts.com/apk