Scaling Down Semantic Leakage: Investigating Associative Bias in Smaller Language Models

TL;DR

This study investigates whether smaller language models exhibit less semantic leakage than larger ones by systematically evaluating models from 500M to 7B parameters using a new color-focused dataset.

Contribution

It introduces a new dataset and evaluation framework to compare semantic leakage across smaller and larger language models, revealing nuanced leakage patterns.

Findings

Smaller models generally show less semantic leakage.

Leakage does not decrease monotonically with model size.

Medium-sized models can sometimes leak more than larger models.

Abstract

Semantic leakage is a phenomenon recently introduced by Gonen et al. (2024). It refers to a situation in which associations learnt from the training data emerge in language model generations in an unexpected and sometimes undesired way. Prior work has focused on leakage in large language models (7B+ parameters). In this study, I use Qwen2.5 model family to explore whether smaller models, ranging from 500M to 7B parameters, demonstrate less semantic leakage due to their limited capacity for capturing complex associations. Building on the previous dataset from Gonen et al. (2024), I introduce a new dataset of color-focused prompts, categorized into specific types of semantic associations, to systematically evaluate the models' performance. Results indicate that smaller models exhibit less semantic leakage overall, although this trend is not strictly linear, with medium-sized models sometimes surpassing larger ones in leaking behavior. The dataset, the model generations, and the evaluation code are publicly available at https://github.com/smilni/semantic_leakage_project.