[NeurIPS 2025] Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness

Overview

This project focuses on our proposed knowledge unlearning framework for LLMs. This codebase provides the official implementation of our framework introduced in the paper.

Environment Setup

Method 1: Using conda

conda env create -f environment.yml
conda activate knowledge_unlearning

Method 2: Using pip

conda create -n knowledge_unlearning python=3.10
conda activate knowledge_unlearning
pip install -r requirements.txt

Project Structure

knowledge_unlearning/
│── src/
│   ├── data/
│   │   ├── knowledge_graph.py  # Handles knowledge graph loading
│   │   ├── text_conversion.py  # Converts KGs to text-based samples            
│   │   ├── yago3-10/ # Select yago3-10 as the knowledge graph
│   ├── models/
│   │   ├── model_loader.py  # Load Model, Tokenizer, and Model Config
│   ├── evaluation/
│   │   ├── evaluator.py  # Query performance of the target LLM and construct subgraph
│   │   ├── llm_evaluator.py  # Evaluate subgraph and target unlearning triples with powerful LLMs
│   ├── main.py  # Main entry point for the project
│   ├── utils/
│   │   ├── utils.py
│   ├── unlearning/
│   │   ├── unlearn.py
│   │   ├── methods/
│   │   │   ├── gradient_ascent.py
│   │   │   ├── random_label.py
│   │   │   ├── negative_preference_optimization.py
│   │   │   ├── ascent_plus_descent.py
│   │   │   ├── scrub.py
│   ├── config/
│   │   ├── config.yml
│   │   ├── config_judge.yml
│   │   ├── config_subgraph.yml
│   │   ├── config_threshold.yml
│   │   ├── config_unlearn.yml
│   │   ├── config_utility.yml
│   ├── README.md

Usage

Unlearn: Load Knowledge Graphs, Unlearn Target Triples over the Target LLM

nohup deepspeed --include localhost:1,2 --master_port=33333 src/main_finetune.py --config config/config_unlearn.yml --unlearn > output.log 2>&1 &

Important Parameters in config_unlearn.yml

• huggingface

  • hf_token: The token associated with your Hugging Face account

• model

  • name: Name of the unlearning model

• unlearn

  • use_QA_unlearning: Set to true to use QA unlearning, false to use Sentence unlearning
  • peft_config: "LoRA" indicates using LoRA-based unlearning; any other value indicates full-parameter unlearning
  • method: Unlearning method to be applied

• output

  • unlearn_model_dir: Path to save the checkpoint

Construct Supporting Subgraph:

nohup python src/main_finetune.py --config config/config_subgraph.yml --subgraph > output.log 2>&1 &

Important Parameters in config_subgraph.yml

• huggingface

  • hf_token: The token associated with your Hugging Face account

• model

  • name: Name of the unlearning model
  • merge_lora: Set to true if the unlearned model was unlearned using LoRA; set to false if it was unlearned using full-parameter unlearning
  • save_merged_model_path: If the unlearned model is trained using LoRA unlearning, this specifies the path to save the merged model after combining the LoRA adapter with the original model
  • checkpoint_path: If the unlearned model is trained using LoRA unlearning, this is the path where the LoRA adapter is saved. If the model is trained with full-parameter unlearning, this is the path where the model parameters are saved

• subgraph

  • k_hop: Number of hops for the Supporting Subgraph
  • unlearn_method: Unlearning method to be applied
  • eval_unlearn_triples: Should be set to true when constructing the Supporting Subgraph
  • score_unlearn_triples: Should be set to false when constructing the Supporting Subgraph

• output

  • subgraph_dir: Directory to save the extracted subgraph

Evaluate Unlearn Triples:

nohup python src/main_finetune.py --config config/config_subgraph.yml --subgraph > output.log 2>&1 &

Similar to Construct Supporting Subgraph, but with the following settings:

• eval_unlearn_triples: set to false
• score_unlearn_triples: set to true

Evaluate Utility Triples:

nohup python src/main_finetune.py --config config/config_utility.yml --utility > output.log 2>&1 &

Important Parameters in config_utility.yml

utility

• run_normal: Specify whether to perform normal utility evaluation.

output

• score_dir: Directory where the results will be saved.

Evaluation with Powerful LLMs:

nohup python src/main_finetune.py --config config/config_judge.yml --judge > output.log 2>&1 &

Important Parameters in config_judger.yml

input

• judge_superficial: Specify whether to score only the unlearn triples.
• judge_superficial_path: Path to the results saved after Evaluate Unlearn Triples.
• judge_subgraph_path: Path to the results saved after Construct Supporting Subgraph.

output

• output_superficial_path: Path to save the scoring results for unlearn triples.
• evaluation_output_file: Path to save the scoring results for supporting subgraphs.

Supported Models

• Qwen/Qwen2.5-7B-Instruct
• meta-llama/Llama-3.1-8B-Instruct

Supported Unlearning Methods

• gradient_ascent
• random_label
• npo
• ascent_plus_descent
• scrub