CleanComedy: Creating Friendly Humor through Generative Techniques
Dmitry Vikhorev, Daria Galimzianova, Svetlana Gorovaia, Elizaveta Zhemchuzhina, Ivan P. Yamshchikov
TL;DR
The paper tackles the challenge of generating humorous text by addressing data toxicity and redundancy through CleanComedy, a bilingual English–Russian humor corpus with partial toxicity filtering and a gold-standard humor annotation subset. It introduces a two-stage training pipeline (Supervised Fine-Tuning followed by LLM Alignment) applied to a LoRA-enhanced LLaMA-3.1-8B, using careful data processing (toxicity filtering, deduplication, and topic modeling) and human evaluation to study humor quality and safety. Key contributions include the public release of CleanComedy, the gold annotations, and empirical evidence that alignment with soft humor scores can reduce toxicity while maintaining comedic relevance across languages. The work advances ethical generative humor research and provides practical insights for deploying humor-capable LLMs in culturally diverse settings, while outlining limitations and directions for larger, more adaptive datasets.
Abstract
Humor generation is a challenging task in natural language processing due to limited resources and the quality of existing datasets. Available humor language resources often suffer from toxicity and duplication, limiting their effectiveness for training robust models. This paper proposes CleanComedy, a specialized, partially annotated toxicity-filtered corpus of English and Russian jokes collected from various sources. We study the effectiveness of our data filtering approach through a survey on humor and toxicity levels in various joke groups. In addition, we study advances in computer humor generation by comparing jokes written by humans with various groups of generative jokes, including our baseline models trained on the CleanComedy datasets.