Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Anatomy of Unlearning: The Dual Impact of Fact Salience and Model Fine-Tuning

Borisiuk Anna, Andrey Savchenko, Alexander Panchenko, Elena Tutubalina

TL;DR

DUAL (Dual Unlearning Evaluation across Training Stages), a benchmark of 28.6k Wikidata-derived triplets annotated with fact popularity using Wikipedia link counts and LLM-based salience scores, shows that pretrained and SFT models respond differently to unlearning.

Abstract

Machine Unlearning (MU) enables Large Language Models (LLMs) to remove unsafe or outdated information. However, existing work assumes that all facts are equally forgettable and largely ignores whether the forgotten knowledge originates from pretraining or supervised fine-tuning (SFT). In this paper, we introduce DUAL (Dual Unlearning Evaluation across Training Stages), a benchmark of 28.6k Wikidata-derived triplets annotated with fact popularity using Wikipedia link counts and LLM-based salience scores. Our experiments show that pretrained and SFT models respond differently to unlearning. An SFT step on the forget data yields smoother forgetting, more stable tuning, and 10-50% higher retention, while direct unlearning on pretrained models remains unstable and prone to relearning or catastrophic forgetting.

Anatomy of Unlearning: The Dual Impact of Fact Salience and Model Fine-Tuning

TL;DR

DUAL (Dual Unlearning Evaluation across Training Stages), a benchmark of 28.6k Wikidata-derived triplets annotated with fact popularity using Wikipedia link counts and LLM-based salience scores, shows that pretrained and SFT models respond differently to unlearning.

Abstract

Machine Unlearning (MU) enables Large Language Models (LLMs) to remove unsafe or outdated information. However, existing work assumes that all facts are equally forgettable and largely ignores whether the forgotten knowledge originates from pretraining or supervised fine-tuning (SFT). In this paper, we introduce DUAL (Dual Unlearning Evaluation across Training Stages), a benchmark of 28.6k Wikidata-derived triplets annotated with fact popularity using Wikipedia link counts and LLM-based salience scores. Our experiments show that pretrained and SFT models respond differently to unlearning. An SFT step on the forget data yields smoother forgetting, more stable tuning, and 10-50% higher retention, while direct unlearning on pretrained models remains unstable and prone to relearning or catastrophic forgetting.
Paper Structure (28 sections, 7 equations, 9 figures, 9 tables)

This paper contains 28 sections, 7 equations, 9 figures, 9 tables.

Table of Contents

  1. Introduction
  2. DUET Benchmark
  3. Data Source.
  4. Filtering and QA Generation.
  5. Popularity Validation.
  6. Popularity-Based Splits.
  7. City Sets.
  8. Experiments
  9. Experimental Setup
  10. Key Findings
  11. Conclusion
  12. Related Work
  13. Machine Unlearning: Definition and Algorithms
  14. Problem Definition
  15. Task Formulation
  16. ...and 13 more sections

Figures (9)

  • Figure 1: Unlearning landscape across fact popularity and model training type. Existing unlearning work does not account for popularity, implicitly assuming all facts are equal. Most studies evaluate forgetting on pretrained or SFT models without contrasting the two. The under-explored quadrant concerns analyzing how unlearning differs between pretrained and SFT models when fact popularity is considered.
  • Figure 2: Overview of the DUET benchmark. (a) Topic distribution across 25 semantic themes (28.6k validated QA pairs), dominated by the places city domain. (b) The filtering and stratification to derive rare and popular forget sets and the retain intersection. (c) Compact retain subsets (retain 500 and retain 1500) drawn from places city, providing resource-efficient yet structurally consistent evaluation settings.
  • Figure 3: City, $n=5\%$. Top: rare; bottom: popular. Left: forget (city_forget_{rare,popular}_5); right: retain (city_fast_retain_500). ROUGE is reported across learning rates. Pretrain is shown as a solid line with circles; SFT as a dashed line with stars. Baselines are solid horizontal lines.
  • Figure 4: City, $n=1\%$. Top: rare; bottom: popular. Left: forget (city_forget_{rare,popular}_1); right: retain (city_fast_retain_500). ROUGE is reported across learning rates. Pretrain is shown as a solid line with circles; SFT as a dashed line with stars. Baselines are solid horizontal lines.
  • Figure 5: City, $n=10\%$. Top: rare; bottom: popular. Left: forget (city_forget_{rare,popular}_1); right: retain (city_fast_retain_500). ROUGE is reported across learning rates. Pretrain is shown as a solid line with circles; SFT as a dashed line with stars. Baselines are solid horizontal lines.
  • ...and 4 more figures