LLMSanitize: Detect Contamination in NLP Data and LLMs

Overview

LLMSanitize is a Python library for investigating contamination between evaluation data and the data used to train or refine language models. It packages a range of published detection approaches, helping NLP researchers and model evaluators compare methods rather than build each one from scratch.

The methods cover both open-data checks, such as string matching and embedding similarity, and closed-data checks that query or inspect a model. It is most useful when evaluating whether benchmark results could be affected by exposure to test examples, and when the evaluator can provide the required model access and infrastructure.

Key Features

• Supports open-data checks using string matching and embedding similarity.
• Includes closed-data approaches based on likelihood, memorization, prompting, and model completion.
• Covers black-box and white-box model access patterns, depending on the method.
• Provides test scripts for running detection methods on datasets such as Hellaswag.
• Can be used through shell scripts or called from Python code.
• Uses a vLLM instance for several methods that require model inference.

Use Cases

• NLP researchers can investigate possible overlap between benchmark examples and model training data before interpreting evaluation results.
• Benchmark maintainers can compare contamination-checking approaches for datasets they publish or curate.
• Model evaluators can assess closed-data contamination when they have access to a model and the method's required inference setup.
• Research teams studying data memorization can run multiple published detection methods within one library.

Project Facts

• Language: Python
• License: Apache-2.0
• Stars: 61
• Forks: 6
• Archived: No

Getting Started

The README specifies Python 3.9 and CUDA 11.8 as the tested setup. Install the package with:

pip install llmsanitize

See the repository README for environment setup, method-specific scripts, and usage details.

Considerations

• The documented setup targets Python 3.9 and CUDA 11.8, and the project notes that it uses vLLM 0.3.3.
• Several methods require launching a vLLM inference server and specifying a port and model name. Other methods may require white-box access to model likelihoods or a model path.
• The available approaches have different data and access assumptions, so a method that fits one evaluation setting may not fit another.
• The repository reports one open issue and was last pushed on 2024-08-13, based on the provided repository metadata.