Sentiment analysis plays a crucial role in many practical applications, but research on sentiment recognition in the unique domain of children’s reading is relatively scarce. The key challenge lies in the distinct linguistic styles and modes of emotional expression in children’s reading, which are often not well captured by existing sentiment analysis models. To address this issue, we first selected five advanced large language models, including Mistral-7b-v0.2, Bloom-7b1, Gemma-7b, LLaMa-2-7b, and LLaMa-2-7b-chat, and applied three different fine-tuning techniques: LoRA, RS-LoRA, and DoRA. Through detailed experimental setups, the performance of each model in the sentiment classification task of children’s reading was compared. The results show that most models significantly improved in performance after applying the RS-LoRA fine-tuning technique, especially in complex sentiment recognition tasks. Notably, the Gemma-7b and Mistral-7b-v0.2 models performed well across all fine-tuning methods, demonstrating high accuracy and F1 scores. However, the performance of the LLaMa-2-7b-chat model slightly declined after applying the DoRA method, indicating that different fine-tuning strategies may have varying effects due to differences in model architectures.