Cendol

Cendol is an open-source collection of instruction-tuned generative language models for Indonesian languages, developed by IndoNLP. It spans decoder-only and encoder-decoder architectures from 300 million to 13 billion parameters, built on bases like mT5 and LLaMA-2, with two tuning tracks: Cendol-Instruct for NLP tasks (sentiment, translation, summarization, QA) and Cendol-Chat for general human-centric conversation.

The project curated the Cendol Collection of roughly 50 million instructions covering 23 tasks and 10 languages, and the paper reports it outperforming existing multilingual and Southeast Asian models on Indonesian evaluations. Everything is released under Apache 2.0, making it the reference open family for Indonesian LLM work.