OPTML-Group/SimNPO-WMDP-zephyr-7b-beta

SimNPO-Unlearned Model on Task "WMDP"

Model Details

• Unlearning:
  • Task: 🤗datasets/cais/wmdp
  • Method: SimNPO
• Origin Model: 🤗HuggingFaceH4/zephyr-7b-beta
• Code Base: github.com/OPTML-Group/Unlearn-Simple
• Research Paper: "Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning"

Unlearning Algorithm

This model uses the SimNPO unlearning algorithm with the following optimization objective: $${\mathrm{\ell }}_{SimNPO}(\mathbf{\theta })={\mathbb{E}}_{(x,y)\in {\mathcal{D}}_f}[-\frac{2}{\beta }\log \sigma (-\frac{\beta }{\mathrm{\mid }y\mathrm{\mid }}\log {\pi }_{\mathbf{\theta }}(y\mathrm{\mid }x)-\gamma )]+\lambda {\mathbb{E}}_{(x,y)\in {\mathcal{D}}_r}[-\log {\pi }_{\mathbf{\theta }}(y\mathrm{\mid }x)]\backslash ell\_\{SimNPO\}(\backslash mathbf\{\backslash theta\})\; =\; \backslash mathbb\{E\}\_\{(x,\; y)\; \backslash in\; \backslash mathcal\{D\}\_f\}\backslash left[-\backslash frac\{2\}\{\backslash beta\}\backslash log\backslash sigma\backslash left(-\backslash frac\{\backslash beta\}\{|y|\}\backslash log\backslash pi\_\{\backslash mathbf\{\backslash theta\}\}(y|x)\; -\; \backslash gamma\backslash right)\backslash right]\; +\; \backslash lambda\; \backslash mathbb\{E\}\_\{(x,\; y)\; \backslash in\; \backslash mathcal\{D\}\_r\}[-\backslash log\backslash pi\_\{\backslash mathbf\{\backslash theta\}\}\; (y|x)]$$ Unlearning hyper-parameters:

• Learning Rate: 4e-6
• beta: 5.5
• lambda: 5.0
• gamma: 0.0

Loading the Model

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("OPTML-Group/SimNPO-WMDP-zephyr-7b-beta", use_flash_attention_2=True, torch_dtype=torch.bfloat16, trust_remote_code=True)

Evaluation Results

	1 - AccBio	1 - AccCyber	MMLU
Origin	0.352	0.608	0.585
NPO	0.581	0.616	0.476
SimNPO	0.584	0.678	0.471

Citation

If you use this model in your research, please cite:

@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}
}

Reporting Issues

Reporting issues with the model: github.com/OPTML-Group/Unlearn-Simple