Sensitivity of Small Language Models to Fine-tuning Data Contamination

TL;DR

The paper systematically investigates how fine-tuning data contamination affects small language models (270M–4B params) deployed in resource-constrained environments. By applying four transformation types (syntactic: character/word reversal; semantic: irrelevant/counterfactual) at contamination levels from $25\%$ to $100\%$, across 23 models from six families, the study reveals a stark asymmetry between syntactic and semantic vulnerabilities and uncovers a 'capability curse' where larger models more readily learn semantic corruptions. It introduces a rigorous evaluation framework combining semantic similarity, lexical metrics, and LLM-based judgments, alongside human agreement assessments, to quantify robustness. The findings challenge assumptions that alignment universally improves robustness and emphasize the need for contamination-aware training protocols and benchmarking to ensure reliable on-device NLP. Overall, the work provides actionable guidance for data curation, model selection, and training strategies to mitigate contamination risks in small, real-world systems.

Abstract

Small Language Models (SLMs) are increasingly being deployed in resource-constrained environments, yet their behavioral robustness to data contamination during instruction tuning remains poorly understood. We systematically investigate the contamination sensitivity of 23 SLMs (270M to 4B parameters) across multiple model families by measuring susceptibility to syntactic and semantic transformation types during instruction tuning: syntactic transformations (character and word reversal) and semantic transformations (irrelevant and counterfactual responses), each applied at contamination levels of 25\%, 50\%, 75\%, and 100\%. Our results reveal fundamental asymmetries in vulnerability patterns: syntactic transformations cause catastrophic performance degradation, with character reversal producing near-complete failure across all models regardless of size or family, while semantic transformations demonstrate distinct threshold behaviors and greater resilience in core linguistic capabilities. Critically, we discover a ``\textit{capability curse}" where larger, more capable models become more susceptible to learning semantic corruptions, effectively following harmful instructions more readily, while our analysis of base versus instruction-tuned variants reveals that alignment provides inconsistent robustness benefits, sometimes even reducing resilience. Our work establishes three core contributions: (1) empirical evidence of SLMs' disproportionate vulnerability to syntactic pattern contamination, (2) identification of asymmetric sensitivity patterns between syntactic and semantic transformations, and (3) systematic evaluation protocols for contamination robustness assessment. These findings have immediate deployment implications, suggesting that current robustness assumptions may not hold for smaller models and highlighting the need for contamination-aware training protocols.

Figures (7)

• Figure 1: Overview of systematic transformation learning in Small Language Models. (I, II) Four transformation types at varying contamination levels (25%-100%) are used to fine tune twenty-three SLMs across six model families (270M-4B parameters). (III) Structural transformations show rapid adoption at 25% contamination, while semantic transformations require higher exposure levels.
• Figure 2: As data contamination increases, SLMs learn to adhere to flawed patterns, causing a decline in task accuracy, semantic similarity, and grammatical correctness.
• Figure 3: Model performance on test data under increasing syntactic data contamination. The heatmaps show model sensitivity (left) and task accuracy (right) for character and word reversal tasks across various contamination levels (25% to 100%).
• Figure 4: Model performance on semantic tasks under increasing data contamination. The figure illustrates instruction-tuned model sensitivity (left) and task accuracy (right) when trained with counterfactual and irrelevant information at different contamination percentages.
• Figure 5: Tokenization of 'Science fiction'