Perkembangan fintech di Indonesia ditandai meningkatnya penggunaan dompet digital seperti DANA dengan ratusan juta pengguna, sehingga mendorong kebutuhan analisis otomatis terhadap ribuan ulasan di Google Play Store. Penelitian ini bertujuan membangun model klasifikasi sentimen menggunakan fine-tuning IndoBERT, model bahasa berbasis Transformer yang dilatih khusus untuk Bahasa Indonesia. Data penelitian diperoleh melalui web scraping 10.000 ulasan terbaru dan diseleksi menjadi 9.470 data valid setelah melalui tahapan pra-pemrosesan teks (case folding, normalisasi, tokenisasi, stopword removal, dan stemming). Pelabelan dilakukan otomatis dengan InSet Lexicon dalam tiga kategori sentimen (positif, netral, negatif) dan dibagi secara stratified ke data latih, validasi, dan uji. Fine-tuning dilakukan pada IndoBERT-base-p1 dengan konfigurasi learning rate 2×10−5, batch size 16, dan 3 epoch. Hasil evaluasi menunjukkan akurasi 92,93% dengan precision, recall, dan F1-score seimbang. Distribusi sentimen mencatat 51,1% positif, 37,3% netral, dan 11,6% negatif, mengindikasikan mayoritas persepsi pengguna terhadap DANA cenderung baik namun masih ada ruang perbaikan. Secara akademis, penelitian ini memvalidasi efektivitas model bahasa mutakhir di domain fintech, sedangkan secara praktis memberi masukan strategis bagi pengembang untuk meningkatkan kualitas produk berbasis data.

The development of fintech in Indonesia, marked by the increasing use of digital wallets like DANA with hundreds of millions of users, has driven the need for automated analysis of thousands of reviews on the Google Play Store. This research aims to build a sentiment classification model using a fine-tuning approach on IndoBERT, a Transformer-based language model specifically trained for the Indonesian language. Research data was obtained through web scraping 10,000 recent reviews, which were filtered into 9,470 valid entries after undergoing text preprocessing stages (case folding, normalization, tokenization, stopword removal, and stemming). Labeling was performed automatically using the InSet Lexicon for three sentiment categories (positive, neutral, negative), and the dataset was then divided using a stratified split into training, validation, and test sets. Fine-tuning was conducted on the IndoBERT-base-p1 model with a configuration of a 2×10−5 learning rate, a batch size of 16, and for 3 epochs. Evaluation results show an accuracy of 92.93% with balanced precision, recall, and F1 score. The sentiment distribution reveals 51.1% positive, 37.3% neutral, and 11.6% negative, indicating that the majority of user perception towards DANA tends to be positive, yet there is still room for improvement. Academically, this research validates the effectiveness of a state-of-the-art language model in the fintech domain, while practically, it provides strategic insights for developers to enhance product quality based on data.