Lius-Cendol-Base-Inst-MT

Lius-Cendol-Base-Inst-MT is an instruction-tuned translation model for Indonesian → Kupang Malay machine translation. The model is part of the Lius model family, developed for low-resource translation using Instructional Linguistic prompting and Continual Instruction Tuning (CIT).

Kupang Malay is a Malay-based creole spoken mainly in East Nusa Tenggara, Indonesia. Despite its broad use as a lingua franca in the region, it remains underrepresented in NLP resources and commercial translation systems. Lius aims to support digital inclusion for Kupang Malay speakers by improving machine translation quality under low-resource conditions.

Model Details

Item	Description
Model name	`joanitolopo/lius-cendol-base-inst-mt`
Model family	Lius
Backbone	`indonlp/cendol-mt5-base-inst`
Architecture	Encoder-decoder / mT5-style sequence-to-sequence model
Parameters	~1.2B
Task	Machine Translation
Direction	Indonesian → Kupang Malay
Source language	Indonesian (`id`)
Target language	Kupang Malay (`mkn`)
Training method	Instructional Linguistic + Continual Instruction Tuning
Developer	Joanito Agili Lopo, Yunita Sari, Guntur Budi Herwanto
Institution	Universitas Gadjah Mada

Model Family

The Lius model family consists of three model sizes:

Model	Parameters	Recommended Use
`joanitolopo/lius-cendol-small-inst-mt`	~350M	Lightweight inference and limited compute
`joanitolopo/lius-cendol-base-inst-mt`	~580M	Balanced performance and efficiency
`joanitolopo/lius-cendol-large-inst-mt`	~1.2B	Best translation quality

Intended Use

This model is intended for:

Translating Indonesian text into Kupang Malay.
Supporting NLP research for Indonesian local languages.
Low-resource machine translation experiments.
Educational and linguistic documentation purposes.
Building prototype translation tools for Kupang Malay.

Out-of-Scope Use

This model is not intended for:

High-stakes legal, medical, financial, or official government translation without human verification.
Translating sensitive content without review by fluent speakers.
Fully replacing native-speaker judgment.
Producing standardized Kupang Malay orthography, since Kupang Malay still has limited orthographic standardization.

Training Data

The model was trained using Indonesian–Kupang Malay parallel data collected and processed from several sources, including:

Bhinneka Korpus
BibleNLP
The Language Archive
Tapaleuk
A bilingual Kupang Malay dictionary based on Kamus Pengantar Bahasa Melayu Kupang

The dataset contains approximately:

Split	Number of Examples
Parallel data	66,521 sentence pairs
Train	53,217 sentence pairs
Test	13,304 sentence pairs
Instruction data	227,172 examples
Experience replay	1,000 examples

The model was trained with instruction-enhanced data generated from several linguistic prompting strategies.

Method

Lius uses an Instructional Linguistic approach, where linguistic information is embedded into the training instructions. Four main instruction types are used:

Context-based instruction Provides relevant Kupang Malay example sentences to help the model understand contextual usage.
Semantic mapping-based instruction Uses bilingual dictionary entries and semantic neighbors to guide lexical meaning transfer.
Phonetic-based instruction Helps the model map phonetically similar words between Indonesian and Kupang Malay.
List-Group-Label-based instruction Groups related Kupang Malay words into semantic categories to enrich vocabulary understanding.

These instruction types are trained using Continual Instruction Tuning (CIT). Instead of training the model only once with a single prompt style, CIT exposes the model to multiple instruction forms for the same translation target, allowing it to learn translation behavior iteratively.

Performance

Evaluation was conducted on Indonesian → Kupang Malay translation using automatic metrics.

Lius Model Comparison

Model	SacreBLEU	chrF++	TER	ROUGE-L
Lius-Small-MT	8.72	26.70	97.96	24.99
Lius-Base-MT	11.63	30.26	92.90	29.75
Lius-Large-MT	12.99	32.01	90.36	31.55

Comparison with Standard Cendol-mT5 Tuning

Model	SacreBLEU	chrF++	TER	ROUGE-L
Cendol-mT5-small-tuned	6.62	23.79	102.50	19.97
Cendol-mT5-base-tuned	8.77	26.91	98.08	25.23
Cendol-mT5-large-tuned	8.97	27.30	97.13	25.94
Lius-Large-MT	12.99	32.01	90.36	31.55

The results show that the Lius instruction-tuned model improves over standard instruction tuning and performs better than several multilingual LLM and NMT baselines in the Indonesian → Kupang Malay translation setting.

Efficiency

Model	Parameters	Storage	Training Time	Inference Time
Lius-Small-MT	350M	1.13 GB	01:09:00	14.21 s
Lius-Base-MT	580M	1.95 GB	03:22:00	22.24 s
Lius-Large-MT	1.2B	3.40 GB	06:35:00	72.35 s

How to Use

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_name = "joanitolopo/lius-cendol-base-inst-mt"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

text = "Saya belum begitu paham dengan maksudnya. Bisakah engkau menerangkannya kembali kepada saya?"

prompt = f"""Teks dalam bahasa Indonesia:
{text}

Terjemahkan ke dalam bahasa Melayu Kupang:"""

inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=128)

outputs = model.generate(
    **inputs,
    max_length=128,
    num_beams=4,
    early_stopping=True
)

translation = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(translation)

Example

Input

Saya belum begitu paham dengan maksudnya. Bisakah engkau menerangkannya kembali kepada saya?

Prompt

Teks dalam bahasa Indonesia:
Saya belum begitu paham dengan maksudnya. Bisakah engkau menerangkannya kembali kepada saya?

Terjemahkan ke dalam bahasa Melayu Kupang:

Output

Beta balóm mangarti deng dia pung maksud bagini. Bosong bisa carita lai sang beta?

Prompt Format

The model was trained using instruction-style prompts. For best results, use a clear translation instruction.

Recommended format:

Teks dalam bahasa Indonesia:
{input_text}

Terjemahkan ke dalam bahasa Melayu Kupang:

Alternative formats may also work, for example:

Apa terjemahan dari teks ini dalam bahasa Melayu Kupang?
Teks: {input_text}

or:

Mohon bantu menerjemahkan teks berikut dari bahasa Indonesia ke bahasa Melayu Kupang.
Teks: {input_text}
Terjemahan:

Limitations

Although Lius improves Indonesian → Kupang Malay translation quality, it still has several limitations:

The model may produce non-standard spelling because Kupang Malay does not yet have a widely standardized orthographic guideline.
The model may hallucinate or add extra information when the input sentence is incomplete or heavily corrupted.
Translation quality may vary depending on dialect, domain, and sentence complexity.
The model may copy some Indonesian words directly, especially named entities, numbers, and formal terms.
Human evaluation shows that native speakers may disagree in judging fluency and adequacy due to dialectal and orthographic variation.
The model should be reviewed by fluent Kupang Malay speakers before deployment in real-world applications.

Bias, Risks, and Ethical Considerations

This model is trained on limited low-resource data. As a result, it may reflect:

Regional dialect bias from the available training sources.
Religious or domain bias from Bible-related data.
Orthographic inconsistency due to the lack of standardized writing conventions.
Limited coverage of modern, technical, or specialized vocabulary.

Users should apply human review when using the model for public-facing or sensitive translation tasks.

Citation

If you use this model, please cite:

@article{lopo2026lius,
  title={Lius: Translation Model Based Instructional Linguistic Using Continual Instruction Tuning in Kupang Malay},
  author={Lopo, Joanito Agili and Sari, Yunita and Herwanto, Guntur Budi},
  journal={arXiv preprint arXiv:2606.11786},
  year={2026}
}

Acknowledgements

This model is developed as part of research on low-resource machine translation for Indonesian local languages. The authors thank the contributors of Kupang Malay linguistic resources, bilingual dictionaries, and open-source Indonesian NLP models that made this work possible.