lm-sys/llm-decontaminator

Code for the paper "Rethinking Benchmark and Contamination for Language Models with Rephrased Samples"

LLM Decontaminator

| Paper | Blog |

In this package, you can use LLM decontaminator to quantify a dataset's rephrased samples relative to a benchmark. Based on the detection results, you can estimate the contamination of rephrased samples in the dataset and remove them from the training set.

Contents

• Install
• Detect
  • Pre-Process
  • End2End
• Real-world dataset
• Dataset and training code
• F1 Score
• Citation

Install

git clone https://github.com/lm-sys/llm-decontaminator.git
cd llm-decontaminator
conda create -n llm-detect python=3.9 -y
conda activate llm-detect
pip install -r requirement.txt

Detect

Pre-Process

Please process the train set and test set into a jsonl format, with each line containing {"text": data}

import json
from datasets import load_dataset

# Load dataset
dataset = load_dataset('bigcode/starcoderdata', data_dir="python", split="train", streaming=True)

# Extract up to 500,000 samples
subset_size = 500000
codes = [sample['content'] for _, sample in zip(range(subset_size), dataset)]

# Write to file
with open("starcoderdata.jsonl", "w") as fout:
    for code in codes:
        fout.write(json.dumps({"text": code}) + "\n")

End2End

# export OPENAI_API_KEY=sk-xxx
# run llm-decontaminator
python3 main.py --train_path ./data/train/CodeAlpaca-20k.jsonl \
    --test_path ./data/test/HumanEval.jsonl \
    --output_path ./data/database/CodeAlpacaDB.jsonl \
    --data-type code \
    --top_k 1

Contamination in Real-world Dataset

Training Set	Benchmark	Train Set Size	Test Set Size	Rephrased Samples	Percentage (%)
The Stack (4G subset)	HumanEval	500k	164	31	18.9
StarCoder-Data (2.4G subset)	HumanEval	500k	164	26	15.9
CodeExercise-Python	HumanEval	27k	164	26	15.9
CodeAlpaca	HumanEval	20k	164	21	12.8
RedPajama-Data-1T (16G subset)	HumanEval	1625k	164	14	8.5
Evol-Instruct-Code	HumanEval	78.3k	164	13	7.9
rossetacode	HumanEval	4.26k	164	4	2.4
MATHInstruct (before Sep 30)	MATH Test	262k	5000	769	15.
MATH Train	MATH Test	7.5k	5000	79	1.6
FLAN CoT	MMLU	184k	14042	76	0.5
WizardLM-Evol-Instruct	MMLU	143k	14042	75	0.5

Dataset and Training Code

Reproduce Llama-rephraser with this document.

F1 Score

Reproduce paper's Table 5 & 6

# MMLU
python3 f1score/mmlu/f1_emb.py
python3 f1score/mmlu/f1_llm.py

# HumanEval
python3 f1score/humaneval/f1_emb.py
python3 f1score/humaneval/f1_llm.py

Table 5:

Table 6:

Citation

Please cite the following paper if you find the code or datasets helpful.

@misc{yang2023rethinking,
      title={Rethinking Benchmark and Contamination for Language Models with Rephrased Samples}, 
      author={Shuo Yang and Wei-Lin Chiang and Lianmin Zheng and Joseph E. Gonzalez and Ion Stoica},
      year={2023},
      eprint={2311.04850},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}