Forgetting and choosing a method: full fine-tuning, LoRA, replay and weight interpolation
Why a fine-tuned model loses abilities it had before, shown with a small experiment you can run (a model that learns task B while task A falls from 98% to 38%), the mitigations and the trade-offs each one buys, how full fine-tuning, LoRA and other methods compare, and a decision table for picking one.
A fine-tuned model usually gets better at the new task. The question people forget to ask is what it got worse at. This lesson shows forgetting happening in a small experiment, then goes through the ways to limit it and the trade-offs each costs, and ends with how to choose between full fine-tuning, LoRA and the other methods.
Pro content
Keep reading with Inference Academy Pro
Unlock the rest of this lesson, including 1 animated diagram, plus every advanced lesson, hands-on lab, the larger playground models and completion certificates.