When Can First-Order Models of Fine-Tuning Bound Forgetting?
First-order LoRA probes bound forgetting only when response coefficients stay within the margin distance.
The paper asks whether measurements at the start of LoRA fine-tuning can bound the chance that protected facts are forgotten on models from 0.6B to 14B parameters. Finite-difference first-order probes predict per-fact margin changes with correlation 0.974-0.998, but direct forgetting predictions fail because updates leave the validated region. Freedman and Azuma first-passage bounds were tested on new runs; the simplified Freedman bound was violated in 14 of 112 conditions. In 43 preregistered confirmatory conditions the complete bound always held, certifying only facts whose response coefficients change by less than their distance to the boundary.
- Finite-difference probes predict margin changes with correlation 0.974-0.998.
- Direct forgetting forecasts fail because updates leave the validated region.
- Simplified Freedman bound failed in 14 of 112 conditions.
- Complete Freedman bound held in all 43 confirmatory conditions.
- Certification requires coefficient change below the fact's margin distance.
From the original article
1 min read at arxiv.orgFine-tuning a language model on new data can make it forget facts that it should keep. We ask whether measurements taken at the start of a fine-tuning run can bound, for each protected fact, the probability that the run makes the model forget it. In LoRA fine-tuning with stochastic gradient descent on models from 0.
By Jianchang Su · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.