The Text mAtching based SequenTial rEcommendation model (TASTE) offers a self-contained recommendation system by mapping items and users into an embedding space and recommending items through text matching. To enhance its performance, this paper proposes TASTE \(^+\) , which adapts language models to sequential recommendation via pretraining. TASTE \(^{+}\) follows the instruction tuning method by designing two pretraining tasks, Masked Item Prediction (MIP) and Next Item Prediction (NIP), which enable language models pretrained on recommendation datasets to better capture text matching signals from user-item interaction sequences. Experimental results on the Yelp and Amazon Product datasets demonstrate that TASTE \(^{+}\) enhances user-item text matching by assigning more balanced attention to both prompt tokens and item IDs. This enables the model to construct more accurate user representations by attending more evenly to recent interaction history. This work demonstrates the effectiveness of language model pretraining in boosting text matching-based recommendation systems.