Finetuning forgetting follows a multiplicative scaling law in model size, finetuning tokens, and injected pretraining fraction, with 1% injection nearly eliminating forgetting.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Scaling Laws for Forgetting during Finetuning with Pretraining Data Injection
Finetuning forgetting follows a multiplicative scaling law in model size, finetuning tokens, and injected pretraining fraction, with 1% injection nearly eliminating forgetting.