Lius-Cendol-Large-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-large-inst-mt
Model family
Lius
Backbone
indonlp/cendol-mt5-large-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
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
python
1from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
23model_name ="joanitolopo/lius-cendol-large-inst-mt"45tokenizer = AutoTokenizer.from_pretrained(model_name)6model = AutoModelForSeq2SeqLM.from_pretrained(model_name)78text ="Saya belum begitu paham dengan maksudnya. Bisakah engkau menerangkannya kembali kepada saya?"910prompt =f"""Teks dalam bahasa Indonesia:
11{text}1213Terjemahkan ke dalam bahasa Melayu Kupang:"""1415inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=128)1617outputs = model.generate(18**inputs,19 max_length=128,20 num_beams=4,21 early_stopping=True22)2324translation = tokenizer.decode(outputs[0], skip_special_tokens=True)25print(translation)
Example
Input
Saya belum begitu paham dengan maksudnya. Bisakah engkau menerangkannya kembali kepada saya?
Prompt
text
1Teks dalam bahasa Indonesia:
2Saya belum begitu paham dengan maksudnya. Bisakah engkau menerangkannya kembali kepada saya?
34Terjemahkan 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:
text
1Teks dalam bahasa Indonesia:
2{input_text}
34Terjemahkan ke dalam bahasa Melayu Kupang:
Alternative formats may also work, for example:
text
1Apa terjemahan dari teks ini dalam bahasa Melayu Kupang?
2Teks: {input_text}
or:
text
1Mohon bantu menerjemahkan teks berikut dari bahasa Indonesia ke bahasa Melayu Kupang.
2Teks: {input_text}
3Terjemahan:
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:
bibtex
1@article{lopo2026lius,
2 title={Lius: Translation Model Based Instructional Linguistic Using Continual Instruction Tuning in Kupang Malay},
3 author={Lopo, Joanito Agili and Sari, Yunita and Herwanto, Guntur Budi},
4 journal={arXiv preprint arXiv:2606.11786},
5 year={2026}
6}
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.
