Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Domains to Instances: Dual-Granularity Data Synthesis for LLM Unlearning

Xiaoyu Xu, Minxin Du, Zitong Li, Zi Liang, Zhibiao Guo, Shiyu Zhang, Peizhao Hu, Qingqing Ye, Haibo Hu

TL;DR

This work formalizes two practical LLM unlearning granularities, domain-level and instance-level, and introduces BiForget, a target-model-guided data-synthesis framework that generates high-quality forget sets without external generators. BiForget employs a two-stage domain-level pipeline (seed-guided synthesis and adversarial probing) and an instance-level rephrasing approach, guided by a semantic-coverage convergence criterion via SimCSE, and evaluates forgetting with a unified metric suite (relevance, diversity, efficiency). Across Harry Potter, WMDP, and TOFU benchmarks, BiForget achieves stronger forgetting with substantially fewer 128-token chunks and better utility preservation than baselines, demonstrating robust, resource-efficient evaluation of unlearning. The approach provides a rigorous foundation for benchmark design and points to scalable extensions for continual and multi-domain unlearning, while acknowledging domain-specific limitations and the need for adaptive prompting.

Abstract

Although machine unlearning is essential for removing private, harmful, or copyrighted content from LLMs, current benchmarks often fail to faithfully represent the true "forgetting scope" learned by the model. We formalize two distinct unlearning granularities, domain-level and instance-level, and propose BiForget, an automated framework for synthesizing high-quality forget sets. Unlike prior work relying on external generators, BiForget exploits the target model per se to elicit data that matches its internal knowledge distribution through seed-guided and adversarial prompting. Our experiments across diverse benchmarks show that it achieves a superior balance of relevance, diversity, and efficiency. Quantitatively, in the Harry Potter domain, it improves relevance by ${\sim}20$ and diversity by ${\sim}$0.05 while halving the total data size compared to SOTAs. Ultimately, it facilitates more robust forgetting and better utility preservation, providing a more rigorous foundation for evaluating LLM unlearning.

From Domains to Instances: Dual-Granularity Data Synthesis for LLM Unlearning

TL;DR

This work formalizes two practical LLM unlearning granularities, domain-level and instance-level, and introduces BiForget, a target-model-guided data-synthesis framework that generates high-quality forget sets without external generators. BiForget employs a two-stage domain-level pipeline (seed-guided synthesis and adversarial probing) and an instance-level rephrasing approach, guided by a semantic-coverage convergence criterion via SimCSE, and evaluates forgetting with a unified metric suite (relevance, diversity, efficiency). Across Harry Potter, WMDP, and TOFU benchmarks, BiForget achieves stronger forgetting with substantially fewer 128-token chunks and better utility preservation than baselines, demonstrating robust, resource-efficient evaluation of unlearning. The approach provides a rigorous foundation for benchmark design and points to scalable extensions for continual and multi-domain unlearning, while acknowledging domain-specific limitations and the need for adaptive prompting.

Abstract

Although machine unlearning is essential for removing private, harmful, or copyrighted content from LLMs, current benchmarks often fail to faithfully represent the true "forgetting scope" learned by the model. We formalize two distinct unlearning granularities, domain-level and instance-level, and propose BiForget, an automated framework for synthesizing high-quality forget sets. Unlike prior work relying on external generators, BiForget exploits the target model per se to elicit data that matches its internal knowledge distribution through seed-guided and adversarial prompting. Our experiments across diverse benchmarks show that it achieves a superior balance of relevance, diversity, and efficiency. Quantitatively, in the Harry Potter domain, it improves relevance by and diversity by 0.05 while halving the total data size compared to SOTAs. Ultimately, it facilitates more robust forgetting and better utility preservation, providing a more rigorous foundation for evaluating LLM unlearning.
Paper Structure (47 sections, 17 equations, 7 figures, 9 tables, 2 algorithms)

This paper contains 47 sections, 17 equations, 7 figures, 9 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Target-Model-Guided Synthesis
  3. Preliminaries and Formulation
  4. LLM Unlearning
  5. Formulating Two Forgetting Scenarios
  6. Domain-level Forgetting.
  7. Instance-level Forgetting.
  8. Methodology
  9. Overview
  10. Domain-level synthesis
  11. Instance-level Synthesis
  12. Evaluation Metrics
  13. Synthesis Analysis
  14. Setup.
  15. Semantic Coverage and Convergence.
  16. ...and 32 more sections

Figures (7)

  • Figure 1: Domain-level vs. Instance-level forgetting
  • Figure 2: BiForget Overview: a target-model-guided synthesis framework for constructing high-quality datasets for domain- and instance-level unlearning, employing seed-guided and adversarial prompts in two stages.
  • Figure 3: Semantic Coverage during synthesis: Cosine similarity rises with # of synthetic samples and finally converges for both domain-level and instance-level.
  • Figure 4: Remote-Clique parameter sensitivity: it stabilizes near $(6, 1000, 1000)$ across points_per_round, $N$, and $M$, indicating stability beyond these values.
  • Figure 5: t-SNE visualization of top-200 chunk embeddings and their centroids for Harry Potter, biosecurity, and cybersecurity: Ours performs best on Harry Potter and biosecurity, but underperforms the Official on cybersecurity.
  • ...and 2 more figures