Review of Grey Box/Black Box Data Contamination Metrics on Open and Commercial Models
Dataset contamination is a problem where benchmarks and tasks used to evaluate the capabilities of Large Language Models (LLMs) have been incorporated into the training dataset of the models. This gives a false sense of performance that can overestimate how these models will function on truly unseen data. This problem becomes worse with commercial LLMs with larger and non-accessible training data, so techniques have been developed to try to measure the degree to which a model is contaminated with a benchmark’s data. To understand the effectiveness of these techniques, particularly when evaluating contamination on coding tasks, we review trends and categorize techniques by the degree of access to the model that is required. The research literature on this topic has reported mixed effectiveness of these techniques, so we select a set of black box (text access only) and grey box (access to model loss/probabilities required) techniques and apply them to both commercial and non-commercial models. We implement these metrics as part of a framework to test the contamination of Python code in LLMs to see to what extent we can replicate the effectiveness (or ineffectiveness) of these contamination detection techniques. Though we find mixed results in the capabilities of these metrics to identify contamination, we do observe evidence that they can identify contamination (broadly) in fine-tuned models when both a baseline and fine-tuned model is present. Additionally, similarity metrics were able to identify between contaminated and uncontaminated data even in situations where the data is distributionally similar (e.g., drawn from the same set of code projects).