This paper proposes a method to align large language models (LLMs) with recommendation knowledge by fine-tuning them with data samples that encode recommendation-specific knowledge. The key idea is to bridge the knowledge gap between LLMs and recommendation systems by generating auxiliary-task data samples that mimic classical operations used in conventional recommender systems, such as masked item modeling (MIM) and Bayesian Personalized Ranking (BPR). These operations are simulated through natural language prompts to generate data samples that encode item correlations and user preferences. The LLMs are then fine-tuned on these auxiliary-task data samples, along with more informative recommendation-task data samples, to inject recommendation-specific knowledge into the models. The proposed method generates recommendation-task data samples that are more informative than existing ones by reducing input/output space complexity and incorporating item titles. The auxiliary-task data samples are generated by simulating MIM, masked language modeling (MLM), and BPR tasks using natural language prompts. These data samples are used to supplement the recommendation-task data samples, enabling the LLMs to better understand the target recommendation domain. Experiments on three real-world datasets (Amazon Toys & Games, Beauty, and Sports & Outdoors) show that the proposed method significantly outperforms conventional and LLM-based baselines in retrieval, ranking, and rating prediction tasks. The method achieves state-of-the-art performance in retrieval, demonstrating its potential for enhancing recommendation quality. The results indicate that the proposed approach effectively aligns LLMs with new recommendation domains by introducing recommendation knowledge through auxiliary-task and recommendation-task data samples. The method is also shown to be effective regardless of the size of the backbone model, and it outperforms existing methods in terms of performance and efficiency.This paper proposes a method to align large language models (LLMs) with recommendation knowledge by fine-tuning them with data samples that encode recommendation-specific knowledge. The key idea is to bridge the knowledge gap between LLMs and recommendation systems by generating auxiliary-task data samples that mimic classical operations used in conventional recommender systems, such as masked item modeling (MIM) and Bayesian Personalized Ranking (BPR). These operations are simulated through natural language prompts to generate data samples that encode item correlations and user preferences. The LLMs are then fine-tuned on these auxiliary-task data samples, along with more informative recommendation-task data samples, to inject recommendation-specific knowledge into the models. The proposed method generates recommendation-task data samples that are more informative than existing ones by reducing input/output space complexity and incorporating item titles. The auxiliary-task data samples are generated by simulating MIM, masked language modeling (MLM), and BPR tasks using natural language prompts. These data samples are used to supplement the recommendation-task data samples, enabling the LLMs to better understand the target recommendation domain. Experiments on three real-world datasets (Amazon Toys & Games, Beauty, and Sports & Outdoors) show that the proposed method significantly outperforms conventional and LLM-based baselines in retrieval, ranking, and rating prediction tasks. The method achieves state-of-the-art performance in retrieval, demonstrating its potential for enhancing recommendation quality. The results indicate that the proposed approach effectively aligns LLMs with new recommendation domains by introducing recommendation knowledge through auxiliary-task and recommendation-task data samples. The method is also shown to be effective regardless of the size of the backbone model, and it outperforms existing methods in terms of performance and efficiency.