Teaching a Small LLM a New Character Erased Its Old Stories: Perplexity Rose From 9 to 156
A tutorial author fine-tuned a 5.9-million-parameter story model, trained on bedtime stories about Momo the monkey, to learn a new character, Zuri the zebra. Training on 300 new Zuri stories took about fifteen seconds of effort but made the model roughly seventeen times worse on its original stories, a failure known as catastrophic forgetting. The author explains that no memories are deleted; the same model weights are repeatedly adjusted toward the new task, so the old capability gradually degrades. To measure this, the author tracks perplexity, where lower is better, on the old stories. The post also describes a remedy being tested, replaying a share of original training windows during fine-tuning, and compares a run without replay against one using 10% replay.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in