apeleg/LKF-baseline-challenger-retain-eval

LKF baseline data for the Challenger disaster topic: retain evaluation

This dataset holds the semantic retain evaluation questions of the LKF baseline comparison data used in the paper "Forget Narrowly, Retain Broadly: Unlearning as an Asymmetric Generalization Problem". It adapts the Challenger disaster topic of the LKF (Lesser-Known Facts) dataset, introduced in "Unlearning That Lasts" (Singh et al., 2025), to the format of our unlearning pipeline.

It is not part of the SUITE benchmark. It exists to reproduce the paper's dataset comparison experiment, in which the same Challenger disaster facts are unlearned by training on LKF data instead of SUITE data.

The three baseline repositories

Repository	Contents
LKF-baseline-challenger-train	training splits
LKF-baseline-challenger-forget-eval	forget evaluation questions with rephrasings
LKF-baseline-challenger-retain-eval (this repository)	semantic retain evaluation questions with rephrasings

Splits and columns

One split, retain_eval_semantic, with 125 questions about topics semantically close to the Challenger disaster (for example the Columbia disaster, the Apollo 1 fire, the Chernobyl disaster, and the Apollo 13 mission). The label column marks the neighboring topic. Each row carries the original question plus 3 paraphrases (q_gemini1 to q_gemini3) and 2 fill-in-the-blank variants (blank_gemini1 to blank_gemini2), used to measure over-forgetting on retained knowledge.

Split	Rows	Columns
retain_eval_semantic	125	question, answer, label, q_gemini1 ... q_gemini3, blank_gemini1 ... blank_gemini2

Usage

The paper's code consumes these datasets through the challenger_baseline topic config:

TOPIC=challenger_baseline bash scripts/suite_unlearn.sh

This experiment runs on Llama only and uses its own hyperparameter defaults (10 epochs, random forget-retain pairing).

Related datasets

The SUITE benchmark itself: apeleg/SUITE and apeleg/SUITE-rephrasings. The full taxonomy and results are on the project page.

Citation

@article{peleg2026forget,
    title   = {Forget Narrowly, Retain Broadly: Unlearning as an Asymmetric Generalization Problem},
    author  = {Peleg, Amit and Singh, Naman Deep and Pearl, Naama and Mohapatra, Bibhabasu and Hein, Matthias},
    journal = {arXiv preprint arXiv:2607.09236},
    year    = {2026}
}

The original LKF dataset:

@article{singh2025unlearning,
    title   = {Unlearning That Lasts: Utility-Preserving, Robust, and Almost Irreversible Forgetting in LLMs},
    author  = {Singh, Naman Deep and M{\"u}ller, Maximilian and Croce, Francesco and Hein, Matthias},
    journal = {arXiv preprint arXiv:2509.02820},
    year    = {2025}
}