Email Augmentation: A Comparison of Fine-tuned Small Language Models

The reliance on outdated datasets poses significant challenges for modern phishing detection systems, as they fail to capture the linguistic diverse and contextual relevance of AI generated email messages. To address this we propose augmenting phishing email datasets using fine-tuned small language models (SLMs) to generate adaptable, privacy-preserving email templates. We evaluate two SLMs—GPT-2 (1.5B parameters) and Gemma3 (1B parameters)—fine-tuned on preprocessed phishing and benign email messages. Our approach combines template generation via structured prompts with preprocessing to anonymise sensitive data and fine-tuning. Results for phishing generation showed that GPT-2 achieves higher alignment with reference templates, while Gemma3 generated more novel and diverse outputs. For benign emails, both models achieved very similar results. The findings highlight the importance of model selection based on use-case priorities, the usage of fine-tuning in addition to prompting to mitigate privacy risks and the viability of SLMs for the generation of email templates.