Muhammad Rafi Fayzulhaq, an undergraduate student in the Informatics Department at Universitas Islam Indonesia (UII), presented his paper at the 2026 International Conference on Applied Artificial Intelligence, Data Engineering and Sciences (ICAIDES 2026) on 12 November 2026. Rafi conducted the research under the supervision of Ahmad Fathan Hidayatullah, S.T., M.Cs., Ph.D. The conference is jointly organized by Universitas Tarumanagara and technically co-sponsored by the IEEE Indonesia Section. This year's hybrid edition carried the theme "Artificial Intelligence for Sustainable and Renewable Engineering Systems," and Rafi presented online in Parallel Session I, Room E.
The paper, "Sentiment Analysis of Indonesian YouTube Comments on Cryptocurrency Using a Fine-Tuned Large Language Model," tackles a practical problem. Discussions about cryptocurrency on YouTube are informal, mix Indonesian slang with English terms, and often include scam warnings, which makes sentiment hard to classify. To address this, the study built a dataset of 4,193 comments, labeled them with a rule-guided large language model, and checked a sample of the labels manually.
The classifier is Qwen2.5-0.5B, adapted with LoRA. It is small enough to run on a standard laptop without a GPU, and comments do not need to be sent to an external service. The study compared three approaches to class imbalance: a baseline, class-weighted loss, and random oversampling. In this run, oversampling gave slightly higher overall accuracy, while weighted loss identified more Negative comments, a trade-off that depends on how the results will be used.
Key takeaway for readers: a small language model can learn sentiment from LLM-assisted labels and run on an ordinary laptop, which makes this approach practical for researchers and communities who cannot send large volumes of comments to external services. The choice of imbalance handling matters, because it changes which kind of error the model makes. The labels are not a human gold standard, however, so the results should be read as a starting point for screening comments, not as a final verdict on any individual comment.








