<![CDATA[The Python ecosystem currently dominates machine learning development, while machine learning libraries in TypeScript remain relatively limited, particularly for deep learning and sequence modeling tasks. This study evaluates the stability of @akhyar11/ml-v1, a TypeScript-native machine learning library, in training recurrent neural network architectures for Indonesian sentiment classification. Rather than focusing solely on identifying the best-performing model, this research positions RNN, LSTM, and GRU as evaluation instruments to assess the library training reliability, consistency, and robustness. Each architecture was trained independently across 10 repeated runs for 30 epochs using the same sentiment classification dataset and experimental configuration. The evaluation focuses on training stability indicators, including accuracy consistency, macro F1-score, weighted F1-score, performance variation across runs, convergence behavior, and the occurrence of training failures such as NaN values or extreme performance drops. The results show that @akhyar11/ml-v1 can run recurrent neural network training processes functionally and relatively stable on the tested configuration. Among the evaluated architectures, LSTM demonstrated the most consistent performance, while RNN showed stable but lower predictive performance. GRU exhibited higher sensitivity across repeated runs, indicating greater variance in training outcomes. These findings suggest that @akhyar11/ml-v1 has practical potential as a non-Python machine learning library for sequence-based natural language processing experiments, although further evaluation on broader datasets, architectures, and training configurations is still required.]]>
Copyrights © 2026
