Fresh human data is what stops models forgetting
A new paper shows that models trained on model-made text lose rare knowledge first, and that a share of fresh data holds the line.
Picture made for ABUREAU.
A paper posted on 4 November asks who pays when models learn from other models' output. Its answer is that someone has to keep supplying fresh material, and that supply is the real constraint.
In small tests on public-domain books, a model trained only on its own output kept 0.20 of its rare words by generation 12. With a tenth fresh text it kept 0.49, with half 0.86. Keeping only the most likely outputs made things worse, not better. Checks that bring in outside information helped.
The author also finds that when the world changes, the fresh share needed rises. The models were small and the setting simple, and the paper says so. The code and per-run results are public.
For people who own data, the point is plain. Data written by people who found things out for themselves is the input that machine-made text cannot replace, and a model that copies its own past needs more of it as the world moves. That is a case for treating well-documented, current human data as something with a price.
