A study in the Arabian Journal for Science and Engineering compared three instruction-tuned models—ALLaM-7B, Qwen2.5-7B, and Llama-3-8B—for generating Hadith-aligned Arabic children’s stories. The models were fine-tuned with QLoRA on a structured Modern Standard Arabic children’s story corpus and given controlled story attributes including age, topic, moral value, setting, tone, and dialogue requirements. Generated stories were used to retrieve supporting Hadith from a corpus of more than 34,000 authenticated narrations via a pipeline using structured query construction, BGE-M3 dense retrieval, cross-encoder re-ranking, and Arabic-aware deduplication. Experts evaluated both stories and retrieved Hadith. Results showed ALLaM-7B achieved the highest overall story quality with a mean score of 3.60/5, followed by Llama-3-8B (3.23) and Qwen2.5-7B (3.03). For Hadith retrieval, ALLaM-7B and Llama-3-8B outputs achieved maximum Hit@3 and MRR scores of 1.000, with nDCG@3 scores of 0.954 and 0.980, respectively. The findings suggest Arabic-centric model design improves story fluency and coherence, and clearer generated morals lead to stronger Hadith retrieval.