mziarehman4353/LRSA-LLM

Stress-Testing Large Language Models (LLMs) against Code-Mixing and Distributional Shifts in Low-Resource NLP # Stress-Testing Large Language Models (LLMs) against Code-Mixing and Distributional Shifts in Low-Resource NLP ## Authors Zia Ul Rehman Zafar Dedi Gunawan Endang Wahyu Pamungkas Widi Widayat Helmi Imaduddin Department of Informatics Engineering Universitas Muhammadiyah Surakarta, Indonesia --- ## Overview This repository accompanies the research study: **Stress-Testing Large Language Models (LLMs) against Code-Mixing and Distributional Shifts in Low-Resource NLP** The study investigates the robustness, generalization capability, and statistical reliability of Large Language Models for Indonesian sentiment analysis under realistic low-resource conditions. The framework evaluates: - Linear SVM (TF-IDF baseline) - IndoBERT - XLM-RoBERTa (XLM-R) across: - Cross-domain distributional shifts - Synthetic noise perturbations - Informal social-media language - Formal news language - Code-mixing scenarios --- ## Dataset The corpus contains approximately 75,000 Indonesian-language sentiment samples collected from: - YouTube comments - Online news articles and comments Domains include: - Politics - Economics - Social issues - Public policy Sentiment labels: - Negative - Neutral - Positive --- ## Experimental Pipeline 1. Data Collection 2. Data Preprocessing 3. Sentiment Modeling 4. Robustness Stress Testing 5. Cross-Domain Evaluation 6. Statistical Significance Testing --- ## Repository Structure models/ results/ figures/ appendix_figures/ data/ notebooks/ src/ docs/ --- ## Citation If you use this repository, please cite the associated publication. --- ## License MIT License ## Suplimentary Materials The transformer models and Dataset are hosted separately on Hugging Face due to their size. ### IndoBERT huggingface.co ### XLM-R huggingface.co ### Dataset huggingface.co ### DOI Archive: doi …