Benchmark-leakage-detection

Benchmark-leakage-detection is a method about data leakage detection. We conducted leakage detection experiments on four benchmark: MMLU, CMMLU, C-Eval and CMB. More details can be found in our paper: Training on the Benchmark Is Not All You Need.

⭐️ Note: The data leakage in the paper refers to the fact that the benchmark test data is included in the training data of LLMs and does not represent artificial cheating. In addition, CMB is a benchmark in the medical field and is not usually targeted for data cleaning by LLM developers.

👑Benchmark leakage leaderboard in LLMs (Continuously updated...)

How to Evaluate on Benchmark-leakage-detection

Data process

First, you need to prepare the data you need to detect, please refer to example_data.json for details. It should be noted that MMLU, CMMLU, and C-Eval have 4 options each, while CMB has 5 options.

{
 'option': {
 'A': '由间充质增生形成', 
 'B': '人胚第4周出现', 
 'C': '相邻鳃弓之间为鳃沟',
 'D': '共5对鳃弓',
 "E": "位于头部两侧"
  },
 'question': '下列有关鳃弓的描述，错误的是'
}

Then use this command to obtain the permutations_data.json.

python data_process.py  --data_dir data_dir --save_dir data

Inference logprobs

Second,use this command to obtain the logprobs.json.

CUDA_VISIBLE_DEVICES=0 python inference_logprobs.py --model_dir model_dir --permutations_data_dir data/permutations_data.json --save_dir data

Alternatively, you can use data parallelism on multiple GPUs.

python3 inference_logprob_parallel.py --model_dir model_dir --permutations_data_dir data/permutations_data.json --save_dir data --world_size 8

Get outlier

Finally,use this command to obtain outlier-(thresholds/max).json.

python get_outlier.py --logprobs_dir data/logprobs.json --permutations_data_dir data/permutations_data.json --save_dir data --method shuffled --permutation_num 24

'permutation_num' represents the factorial of the number of your options. You can specify 'shuffled' or 'not_shuffled' as the parameter for the 'method', where 'not_shuffled' and 'shuffled' represent 'Scenario a' and 'Scenario b', respectively.

Citation

@article{ni2024training,
  title={Training on the Benchmark Is Not All You Need},
  author={Ni, Shiwen and Kong, Xiangtao and Li, Chengming and Hu, Xiping and Xu, Ruifeng and Zhu, Jia and Yang, Min},
  journal={arXiv preprint arXiv:2409.01790},
  year={2024}
}

TODO

• add paper link