Views
No views yet
SimNPO unlearning algorithm with the following optimization objective:
$$\ell_{SimNPO}(\mathbf{\theta}) = \mathbb{E}{(x, y) \in \mathcal{D}f}\left[-\frac{2}{\beta}\log\sigma\left(-\frac{\beta}{|y|}\log\pi{\mathbf{\theta}}(y|x) - \gamma\right)\right] + \lambda \mathbb{E}{(x, y) \in \mathcal{D}r}[-\log\pi{\mathbf{\theta}} (y|x)]$$
Unlearning hyper-parameters:1e-50.71.00.01import torch
2from transformers import AutoModelForCausalLM, AutoTokenizer
3
4model = AutoModelForCausalLM.from_pretrained("OPTML-Group/SimNPO-MUSE-Books-iclm-7b", torch_dtype=torch.bfloat16, device_map='auto')| VerbMem Df | KnowMem Df | PrivLeak | KnowMem Dr | |
|---|---|---|---|---|
| Origin | 99.56 | 58.32 | -56.32 | 67.01 |
| Retrain | 14.30 | 28.90 | 0.00 | 74.50 |
| NPO | 0.00 | 0.00 | -31.17 | 23.71 |
| SimNPO | 0.00 | 0.00 | -19.82 | 48.27 |
@article{fan2024simplicity,
title={Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning},
author={Fan, Chongyu and Liu, Jiancheng and Lin, Licong and Jia, Jinghan and Zhang, Ruiqi and Mei, Song and Liu, Sijia},
journal={arXiv preprint arXiv:2410.07163},
year={2024}
}