Fine-Tuning Compact LLMs for Controllable Children's Story Generation

Researchers have developed a method to generate children's English reading stories using compact 8B-parameter large language models (LLMs) fine-tuned on expert-designed curricula. The study, published on arXiv, addresses two key issues: stories generated by LLMs are often too difficult for children, and the high operational cost of large models limits their educational use. By fine-tuning three 8B LLMs on stories from GPT-4o and Llama 3.3 70B, the team achieved better control over reading difficulty and error patterns compared to zero-shot generation from larger models. The approach prioritizes controllability over scale, allowing educators to target specific reading levels with affordable models. Evaluation results showed that fine-tuned 8B models outperformed larger zero-shot models on difficulty-related metrics.