Execution Steps

This section provides step-by-step instructions for executing the CLUE project using Edge Pruning methodology. The paper is “CLUE: Conflict-guided Localization for LLM Unlearning Framework” accepted by ICLR 2026.

Step 1: Dataset Creation

Execute the dataset creation script using Python:

python Edge-Pruning/get_wmdp.py

Additional datasets can be obtained from the Edge-Pruning repository.

Note: If you encounter insufficient GPU memory, you can modify the max_token parameter to control the token length of the generated dataset. However, this may lead to unexpected deviations in the final results.

Step 2: Circuit Generation

Execute the pruning scripts to generate Noising and Denoising circuits:

python Edge-Pruning/src/prune/fmistral_cyber.py
python Edge-Pruning/src/prune/fmistral_cyber_or.py

Please execute the corresponding Python files based on your specific dataset requirements.

Step 3: Edge Refinement

Generate refined edges.json files using the visualization script:

python Edge-Pruning/src/modeling/vis_mistral.py

Note: Noising and denoising processes will produce two different JSON files.

Step 4: Logical Circuit Generation

Extract circuits containing conjunction and disjunction operations:

python Edge-Pruning/src/solver/get_logical_edge.py

Step 5: Conflict Analysis

Identify forget nodes and conflict nodes:

python Edge-Pruning/src/solver/conflict.py

Step 6: Mask Generation

Generate masks for the corresponding nodes:

python Edge-Pruning/src/solver/mask_generation.py

Step 7: Model Unlearning

Perform the unlearning process:

python src/exec/unlearn_model_conlict.py

BIB

@inproceedings{
Chen2026clue,
title={{CLUE}: Conflict-guided Localization for {LLM} Unlearning Framework},
author={Chen, Hang and Zhu, Jiaying and Yang, Xinyu and Wang, Wenya},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=jtRYvazBWv}
}