📄 Paper
LastingBench: Defend Benchmarks Against Knowledge Leakage.
Welcome to the repository for the research paper: "LastingBench: Defend Benchmarks Against Knowledge Leakage." This project addresses the growing concern about large language models (LLMs) "cheating" on standard Question Answering (QA) benchmarks by memorizing task-specific data, which undermines the validity of benchmark evaluations as they no longer reflect genuine model capabilities but instead the effects of data leakage.
Project Overview
Overview LastingBench introduces a novel framework designed to continuously reinforce and safeguard existing benchmarks against knowledge leakage. The project aims to:
Detect knowledge leakage through context and question perturbation techniquesRewrite leaked content to counterfactual alternatives that disrupt memorization while preserving the benchmark's original evaluative intentEvaluate model responses to contextual evidence and reasoning patternsProvide practical solutions to ensure benchmark robustness over time, promoting fairer and more interpretable evaluations of LLMs
Installation
Clone the repository:
git clone https://github.com/Seriousss/lastingbench
Create and activate conda environment:
1 conda create -n lastingbench python = 3.12
2 conda activate lastingbench
Install dependencies:
pip install -r requirements.txt
Set up environment variables:
1 export OPENAI_BASE_URL = "your-api-base-url"
2 export OPENAI_API_KEY = "your-api-key"
3 export CUDA_VISIBLE_DEVICES = "0,1,2,3" # Adjust based on your GPU setup
Usage
LastingBench provides three main functionalities: Detection , Rewrite , and Training Comparision .
🔍 Detection
Detect knowledge leakage through various perturbation techniques.
1. Context Leakage Detection
Evaluate models using exact-match scoring on benchmark datasets:
1 # Using vLLM for most models
2 python -m detect.contextleakage --hf_model "Qwen/Qwen2.5-7B-Instruct" \
3 --dataset_subset "hotpotqa" --cuda_devices "0,1"
4
5 # Using Transformers for Qwen3 models
6 python -m detect.contextleakage --hf_model "Qwen/Qwen3-8B" \
7 --is_qwen3 --max_new_tokens 30
8
9 python -m detect.contextleakage_api --model "deepseek-r1" --dataset_subset "hotpotqa"
2. Question Perturbation Detection
Rephrase questions to opposite meanings and test model consistency:
1 # Using OpenAI API
2 python -m detect.question_rephrase_answer_api \
3 --model_name "gpt-4o" --dataset_subset "2wikimqa" \
4 --rephrase_type "opposite" --sample_count 100
5
6 # Using local vLLM models
7 python -m detect.question_rephrase_answer_vllm \
8 --model_name "Qwen/Qwen2.5-7B-Instruct" --dataset_subset "hotpotqa" --rephrase_type "similar"
9
10 # Using Qwen3 with Transformers
11 python -m detect.question_rephrase_answer_qwen3 \
12 --model_name "Qwen/Qwen3-8B" --dataset_subset "2wikimqa"
✏️ Rewrite
Generate counterfactual answers and rewrite leaked evidence to create robust benchmarks.
`
1. Evidence Finding and Counterfactual Rewriting Pipeline
Run the complete finding and rewriting pipeline:
1
2 # Specify custom output file and dataset
3 python main_gpu.py --output custom_output.jsonl \
4 --dataset_subset "hotpotqa" --start_idx 0 --max_samples 100
5
Convert and merge JSONL files with question-answer mappings:
1 # Merge single mapping file with original dataset
2 python utils/convert.py original.jsonl revised.jsonl custom_output.jsonl
3
The original and revised dataset can be found under the data folder.
2. Random Answer Rewriting
Create random alternatives to disrupt memorization:
1 # Specify custom output file and dataset
2 python random_alternative_answer.py --output random_hotpot.jsonl \
3 --dataset_subset "hotpotqa" --start_idx 0 --max_samples 50
4
🚀Dataset evaluations on model inference and training
1. Model Inference Evaluation
Comprehensive evaluation on original and revised benchmarks:
1 # Transformers-based evaluation
2 python -m eval.evaluation -i data/hotpotqa.jsonl -model "Qwen/Qwen3-8B" -k 40 -t 0.5
3
4 # API-based evaluation
5 python -m eval.eval_with_api.py --input data/hotpotqa_antifact.jsonl \
6 --model "deepseek-r1" --max_tokens 30 --temperature 0.5
2. Model training Evaluation
Compare training dynamics between original and rewritten datasets:
The training loss data can be found under training_result .
To repoduce the picture in our paper:
1 python utils/draw.py training_result/training_loss_qwen38.csv training_result/training_loss_antifact_qwen38.csv \
2 --title "Original vs Rewritten Training Loss"
📊 Utility Functions
Additional tools for analysis and metrics:
Metrics Calculation : F1 scores, EM scores, and custom evaluation metricsDocument Retrieval : BM25-based retrieval for evidence analysis
All scripts support various parameters for customization. Use --help with any script to see available options.
