indonesian-nlp

Transformer Encoder for Indonesian intent classification — implementing Positional Encoding & Multi-Head Self-Attention from scratch, compared against fine-tuned BERT via 5-fold cross-validation. Includes overfitting investigation, early stopping, and FastAPI serving.

Daily catalog of new Hugging Face models in vision, speech and Indonesian LLMs, with noise filtering, CSV catalog and daily reports, auto-committed by GitHub Actions.

IdSarcasm: Benchmarking and Evaluating Language Models for Indonesian Sarcasm Detection (Suhartono et al., 2024)

GPT-2 text generation lab: training a code-completion model from scratch (CodeParrot) and serving it via FastAPI, plus fine-tuning a pretrained Indonesian GPT-2. Results are documented honestly, including undertraining and overfitting findings — not just the wins.

A Python NLP project to detect social media hate speech in Indonesian language.

Local evaluation of Qwen2.5 GGUF models (llama-cpp-python): accuracy benchmark on Indonesian questions and OCR-to-JSON structured extraction, with Wilson intervals and McNemar tests.

A curated list of NLP resources for Bahasa Indonesia and Indonesia's regional languages

Intent classification on Indonesian text using TF-IDF + Logistic Regression. 100% accuracy confirmed via 5-fold cross-validation, then stress-tested with edge cases that reveal what that accuracy actually does (and doesn't) mean.

Indonesian Text-to-SQL benchmark for LLM evaluation with 30 scenarios across two open databases.

A simple FAQ search engine built with TF-IDF and Sastrawi (Indonesian stemming). Includes an audit that found Euclidean distance let an off-topic question nearly pass as a match — fixed by switching to cosine similarity.

Offline (EM/F1/BLEU/ROUGE/perplexity/semantic-similarity) and online (FastAPI + llama.cpp + MySQL cache + Streamlit dashboard) evaluation metrics for LLM applications.

QLoRA fine-tuning of Llama 3.1 8B on a translated Indonesian instruction dataset using Unsloth, with before/after evaluation, GGUF export, and a size check against common 4-bit compression claims.

IndoELECTRA: Pre-Trained Language Model for Indonesian Language Understanding

Research code for our ICIMCIS 2026 paper: distant supervision with heuristic-augmented denoising for crypto NER on Indonesian Twitter using IndoBERTweet-CRF

Multi-label hate speech & abusive language detection for Indonesian tweets using deep learning

NLP dashboard that turns TikTok comments about Indonesia's Makan Bergizi Gratis (MBG) program into sentiment, public issues, and risk signals.

A machine learning system for classifying Indonesian text sentiment (positive/negative/neutral) using XGBoost (TF-IDF) and transformer-based NLP, trained on 30K MyPertamina app reviews.

Quantization research for compiler-verified LLM systems. Three arms: (A) Phase D IntLLM 1.58-bit ternary, 3 gates PASS; (B) v3.1 adaptive KV cache quant, 7/9 cells; (C) Phase E bilingual ID+EN, corpus v1.0 25.67B tokens, E2.4 honest negative result. In-kernel via FajarOS Nova. Apache 2.0. Made in Indonesia.

Social media analytics course materials: text mining, topic modelling, data collection (Twitter/X API & Twint), stance detection, user profiling (gender & occupation), and network analysis with PageRank.

Indonesian Gojek review sentiment analysis comparing TF-IDF Logistic Regression, SVM, and a dense neural network.