Accurate identification of antibiotic resistance genes (ARGs) is essential for mitigating the spread of antimicrobial resistance. This study introduces a framework that leverages the Nucleotide Transformer (NT) model to predict ARGs in a hybrid dataset consisting of both annotated and non-annotated genes. The NT model was trained using a two-round screened dataset to improve prediction accuracy. The experiment processed non-annotated genomic sequences from NCBI using two distinct methods: sliding window segmentation and CDS annotation with the Prodigal_test tool. These processed sequences were subsequently analyzed by the fine-tuned NT model to predict their antibiotic resistance values. This comparison aimed to improve the real-world predictive performance of the fine-tuned NT model. Predictions from the fine-tuned NT model were cross-validated with BLAST results to assess their feasibility and reliability for non-annotated genes.The dual-track framework (NT prediction + BLAST validation) significantly enhances the reliability of ARG identification and facilitates iterative updates to antibiotic resistance gene databases.