Neighboring Perturbations of Knowledge Editing on Large Language Models

Authors: Jun-Yu Ma, Zhen-Hua Ling, Ningyu Zhang, Jia-Chen Gu

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate the effectiveness of APP coupling with four editing methods on four LLMs.
Researcher Affiliation Academia 1NERC-SLIP, University of Science and Technology of China 2Zhejiang University 3University of California, Los Angeles.
Pseudocode No The paper describes the proposed method and its objectives using mathematical equations and textual explanations, but it does not include a formal pseudocode block or algorithm.
Open Source Code Yes The code and data are available at https://github.com/mjy1111/PEAK.
Open Datasets Yes To evaluate the additivity of edited models, a benchmark dubbed as Perturbation Evaluation of Appending Knowledge (PEAK) is constructed... The code and data are available at https://github.com/mjy1111/PEAK.
Dataset Splits Yes To set the hyperparameters, we additionally created a small validation set.
Hardware Specification No The paper mentions the large language models used (e.g., GPT-2 XL, LLaMA-2) and alludes to 'limited computing resources', but it does not specify any particular hardware like GPU models, CPU types, or memory specifications.
Software Dependencies No The paper mentions the use of specific LLMs (e.g., GPT-2 XL, GPT-J, LLaMA-2) and frameworks for editing (e.g., Easy Edit), but it does not provide specific version numbers for underlying software libraries like Python, PyTorch, or CUDA.
Experiment Setup Yes Appendix B.2, 'Hyperparameters for APP', states: 'Table 8 shows the details of hyperparameters set for different LLMs. Besides, the margin M is set to 2.' Table 8 then lists specific alpha, beta, and gamma values for each method and LLM.