kixx/LastingBench

📄 Paper

📥 Download the PDF

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

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 techniques
• Rewrite leaked content to counterfactual alternatives that disrupt memorization while preserving the benchmark's original evaluative intent
• Evaluate model responses to contextual evidence and reasoning patterns
• Provide practical solutions to ensure benchmark robustness over time, promoting fairer and more interpretable evaluations of LLMs

Installation

1. Clone the repository:

git clone https://github.com/Seriousss/lastingbench

2. Create and activate conda environment:

conda create -n lastingbench python=3.12
conda activate lastingbench

3. Install dependencies:

pip install -r requirements.txt

4. Set up environment variables:

export OPENAI_BASE_URL="your-api-base-url"
export OPENAI_API_KEY="your-api-key"
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:

# Using vLLM for most models
python -m detect.contextleakage --hf_model "Qwen/Qwen2.5-7B-Instruct" \
    --dataset_subset "hotpotqa" --cuda_devices "0,1"

# Using Transformers for Qwen3 models  
python -m detect.contextleakage --hf_model "Qwen/Qwen3-8B" \
    --is_qwen3 --max_new_tokens 30

python -m detect.contextleakage_api --model "deepseek-r1" --dataset_subset "hotpotqa"

2. Question Perturbation Detection

Rephrase questions to opposite meanings and test model consistency:

# Using OpenAI API
python -m detect.question_rephrase_answer_api \
    --model_name "gpt-4o" --dataset_subset "2wikimqa" \
    --rephrase_type "opposite" --sample_count 100

# Using local vLLM models
python -m detect.question_rephrase_answer_vllm \
    --model_name "Qwen/Qwen2.5-7B-Instruct" --dataset_subset "hotpotqa" --rephrase_type "similar"

# Using Qwen3 with Transformers
python -m detect.question_rephrase_answer_qwen3 \
    --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:

# Specify custom output file and dataset
python main_gpu.py --output custom_output.jsonl \
    --dataset_subset "hotpotqa" --start_idx 0 --max_samples 100

Convert and merge JSONL files with question-answer mappings:

# Merge single mapping file with original dataset
python utils/convert.py original.jsonl revised.jsonl custom_output.jsonl

The original and revised dataset can be found under the data folder.

2. Random Answer Rewriting

Create random alternatives to disrupt memorization:

# Specify custom output file and dataset
python random_alternative_answer.py --output random_hotpot.jsonl \
    --dataset_subset "hotpotqa" --start_idx 0 --max_samples 50

🚀Dataset evaluations on model inference and training

1. Model Inference Evaluation

Comprehensive evaluation on original and revised benchmarks:

# Transformers-based evaluation
python -m eval.evaluation -i data/hotpotqa.jsonl -model "Qwen/Qwen3-8B" -k 40 -t 0.5

# API-based evaluation  
python -m eval.eval_with_api.py --input data/hotpotqa_antifact.jsonl \
    --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:

python utils/draw.py training_result/training_loss_qwen38.csv training_result/training_loss_antifact_qwen38.csv \
    --title "Original vs Rewritten Training Loss"

📊 Utility Functions

Additional tools for analysis and metrics:

• Metrics Calculation: F1 scores, EM scores, and custom evaluation metrics
• Document Retrieval: BM25-based retrieval for evidence analysis

All scripts support various parameters for customization. Use --help with any script to see available options.