TLDR: A new study investigates dataset contamination in Large Language Models (LLMs) when interacting with public tabular datasets like Adult Income and Titanic. Researchers found that LLMs do not memorize specific data entries (syntactic contamination) but instead exhibit strong semantic familiarity with well-known datasets. This means models leverage prior knowledge about feature meanings and contextual associations, which can artificially inflate their performance on reasoning tasks. The findings emphasize the need for contamination-aware evaluation protocols to accurately assess LLM capabilities.
Large Language Models (LLMs) are increasingly being used for tasks that involve reasoning over structured data. However, a critical challenge in evaluating these models is the issue of dataset contamination. This refers to the unintentional inclusion of benchmark data within the LLM’s training corpus, which can make it seem like the model is reasoning when it might just be recalling information it has already seen.
A recent study by Matteo Silvestri, Flavio Giorgi, Fabrizio Silvestri, and Gabriele Tolomei from Sapienza University of Rome delves into this problem, specifically focusing on widely used public tabular datasets. Their research, titled Evaluating Latent Knowledge of Public Tabular Datasets in Large Language Models, investigates whether LLMs possess prior knowledge of benchmarks like Adult Income and Titanic, and how this affects their performance.
Understanding Data Contamination
The researchers define dataset contamination as the presence of examples, labels, or metadata from evaluation datasets within a model’s pretraining data. They distinguish between two main types:
-
Syntactic Contamination: This occurs when exact or near-exact entries from a dataset are included in the training data. For instance, if rows from the “Adult Income” dataset are directly memorized because they appeared in online tutorials.
-
Semantic Contamination: This is more subtle. It happens when the model is exposed to descriptions, summaries, or derived representations of the data, allowing it to learn conceptual patterns or correlations even without seeing the exact text. For example, understanding that “people earning over 50K in the Adult dataset are often men in tech jobs.”
These distinctions are vital because both forms can artificially inflate a model’s performance, making it difficult to assess its true reasoning abilities.
Why Tabular Data is a Special Case
Tabular datasets, unlike free-form text, are structured with columns (features) and rows (instances). While not inherently linguistic, many popular tabular benchmarks are widely available online in tutorials, GitHub repositories, and academic papers. Crucially, these datasets often contain semantic cues, such as meaningful feature names (e.g., “age,” “occupation,” “fare”) or interpretable categorical values. These cues can allow an LLM to leverage prior knowledge, even if it hasn’t memorized specific data entries.
The Research Approach
To investigate contamination, the study employed a systematic experimental framework using two main tasks:
-
Completion Task: LLMs were asked to complete masked attributes (20% of features) in a record, choosing the correct value from five candidates. This task aimed to detect syntactic contamination, or memorized associations between features.
-
Existence Task: Models were presented with five versions of a record (one genuine, four perturbed) and asked to identify the original. This tested for memorization of exact data tuples or high-level structural familiarity.
To differentiate contamination from genuine reasoning, the researchers also created modified versions of datasets:
-
Synthetic Marginal Dataset: Preserved low-level statistical properties but removed inter-feature dependencies, neutralizing syntactic memorization.
-
Obfuscated Dataset: Relabeled categorical attributes with abstract symbols (e.g., c01, c02), disrupting domain-level semantics to prevent models from using contextual knowledge.
They evaluated three families of open-weight LLMs (llama_8B, mistral_7B, and qwen models up to 32B parameters) in a zero-shot setting across a diverse set of semantic (e.g., Adult, Titanic, Credit) and non-semantic (e.g., MAGIC Gamma Telescope, Synthetic Binary Classification) datasets.
Key Findings: Semantic Familiarity, Not Memorization
The experiments revealed a clear pattern: contamination effects emerged exclusively in the completion task, and only for the three most widely known semantic datasets: Adult Income, Titanic, and Default of Credit Card Clients. In these cases, all evaluated models achieved accuracies significantly above a random guess (which was approximately 20%). This suggests that the models likely possess a semantic-level familiarity with these datasets, possibly due to indirect exposure through publicly available resources and tutorials.
Crucially, the existence task, which directly probed for the memorization of specific records, was consistently failed across all models and datasets. This indicates that the LLMs examined (up to 32B parameters) do not exhibit clear syntactic contamination or verbatim memorization of data instances.
The findings suggest that LLMs “do not remember data, but they remember meaning.” Their apparent competence on tabular reasoning tasks, particularly for well-known datasets, seems to stem from leveraging contextual and conceptual knowledge acquired during pretraining, enabling them to infer plausible values when dataset semantics are familiar. This semantic familiarity, while not direct data leakage, can still artificially inflate performance on widely discussed benchmarks.
Implications for LLM Evaluation
This research highlights a significant threat to the validity of LLM evaluations. When a model performs well on a benchmark it has implicitly seen, the result may not reflect genuine reasoning or generalization but rather the retrieval of known patterns. This can undermine claims about interpretability, causal reasoning, and zero-shot learning.
The study urges the community to adopt contamination-aware evaluation protocols for future benchmarks. This includes using less disseminated or synthetic datasets to ensure that measured performance truly reflects an LLM’s reasoning ability rather than latent semantic contamination.
Also Read:
- Unveiling Hidden Data: How Alignment Information Leaks from Open Language Models
- Unveiling LLM Challenges with Dynamic Information: The evolveQA Benchmark
Looking Ahead
While informative, the study acknowledges limitations, such as the focus on models up to 32B parameters and a limited scope of datasets. Future work aims to evaluate larger LLMs and categorize datasets by their empirically observed contamination levels to further understand how contamination impacts reasoning and generalization.
