This paper presents our system for the SemEval-2023 Task 4, which aims to identify human values behind arguments by classifying whether or not an argument draws on a specific category. Our approach leverages a second-phase pre-training method to adapt a RoBERTa Language Model (LM) and tackles the problem using a One-Versus-All strategy. Final predictions are determined by a majority voting module that combines the outputs of an ensemble of three sets of per-label models. We conducted experiments to evaluate the impact of different pre-trained LMs on the task, comparing their performance in both pre-trained and task-adapted settings. Our findings show that fine-tuning the RoBERTa LM on the task-specific dataset improves its performance, outperforming the best-performing baseline BERT approach. Overall, our approach achieved a macro-F1 score of 0.47 on the official test set, demonstrating its potential in identifying human values behind arguments.