小模型更少泄露语义关联，但中等规模模型例外
Scaling Down Semantic Leakage: Investigating Associative Bias in Smaller Language Models
- 用新构建的彩色提示数据集，系统测试500M至7B参数模型的语义泄露
- 小模型整体泄露程度更低，但500M~3B模型间存在非线性波动
- 适合关注模型隐私与可控生成的研究者参考
语义泄露是Gonen等人（2024）提出的新现象，指语言模型在生成过程中意外呈现训练数据中的关联关系。先前研究集中于70亿参数以上的大模型。本研究采用Qwen2.5系列模型，覆盖5亿至70亿参数范围，探究小模型是否因容量限制而减少语义泄露。基于Gonen等人的数据集，构建了聚焦颜色的新型提示数据集，按语义关联类型分类，用于系统评估模型表现。结果表明，总体上小模型泄露程度较低，但该趋势并非严格线性；部分中等规模模型（如30亿参数）甚至表现出比大模型更强的泄露行为。相关数据集、生成结果与评估代码已公开于https://github.com/smilni/semantic_leakage_project。
原文摘要 · Abstract (English)
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.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。