SEA-LION is a collection of Large Language Models (LLMs) which have been pretrained and fine-tuned for the Southeast Asia (SEA) region.
Leveraging the advanced encoder-only ModernBERT architecture combined with the Gemma 3 SentencePiece tokenizer, this specific model achieves highly efficient and culturally nuanced text processing with improved tokenization fertility and compression rates for complex regional scripts, enabling it to handle longer context windows and cross-lingual tasks with greater computational efficiency. Using the SEA-LION-ModernBERT-300M-checkpoints as the base model, this model is contrastively pre-trained on 245 million text pairs (EN-EN and EN-SEA) and finally instruction-tuned using a diverse dataset of 13 million text pairs (spanning EN-EN, CN-CN, EN-SEA, and SEA-SEA) to create the final instruction-tuned model.
Citation
BibTeX:
@misc{limkonchotiwat2026seaembedding,
title = {SEA-Embedding: Open and Reproducible Text Embeddings for Southeast Asia},
author = {Limkonchotiwat, Peerat and Ng, Raymond and Nutanong, Sarana and Ngui, Jian Gang},
year = {2026},
eprint = {2606.03027},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2606.03027}
}
Model Details
Model Description
The SEA-LION-ModernBERT-Embedding-300M model is built on the ModernBERT-base architecture and has a vocabulary size of 262K.
For tokenization, the model employs our custom Gemma 3 tokenizer, which has excellent performance for SEA languages, ensuring optimal model performance.
Repository: We are releasing the weights for every training stage to promote transparency and research and to support a wide range of downstream applications. The SEA-LION-ModernBERT-Embedding-300M model can be accessed at the Link to HF Repo.
Uses
This model card details one of the variants available within this collection.
- Continued Pre-Training (CPT) - Fine-tuning for downstream tasks (e.g., sentiment analysis, classification).
Bias, Risks, and Limitations
The model was not tested for robustness against adversarial usage. It is important for users to be aware that our model exhibits certain limitations that warrant consideration. Users should also exercise caution in continue-implementing and validating the model's responses due to the potential inconsistencies.
Recommendations
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.
This model was tuned using a multi-stage training pipeline with the following datasets:
Contrastive Pre-training: 245 million text pairs (EN-EN and EN-SEA) to enhance cross-lingual alignment.
Fine-tuning: 13 million diverse text pairs (spanning EN-EN, CN-CN, EN-SEA, and SEA-SEA) to create the final fine-tuned model.
Language
Percentage
EN-EN
20%
CN-CN
20%
EN-SEA
10%
SEA-SEA
50%
Training Procedure
Preprocessing
Following the foundational training, the model's cross-lingual alignment was substantially enhanced by undergoing contrastive pre-training utilising 245 million text pairs, specifically focusing on English-to-English and English-to-Southeast Asian language mappings (EN-EN and EN-SEA). Finally, to ensure the model could effectively follow user instructions and handle complex interactions, it was instruction-tuned using a diverse dataset of 13 million text pairs spanning EN-EN, CN-CN, EN-SEA, and SEA-SEA, culminating in the final highly capable instruction-tuned model.
Evaluation
Testing Data, Factors & Metrics
Testing Data
The model is evaluated across three primary benchmark suites to provide a comprehensive assessment of embedding quality across Southeast Asian, Chinese, and English contexts:
SEA-BED (Southeast Asia Embedding Benchmark): The primary testing suite, consisting of 169 datasets across 10 Southeast Asian languages (Burmese, Filipino, Indonesian, Khmer, Malay, Lao, Tamil, Tetum, Thai, and Vietnamese). Notably, 71% of these datasets are native-authored or human-curated to preserve regional linguistic properties.
CMTEB (Chinese Massive Text Embedding Benchmark): A specialised subset of MTEB focused on Chinese language tasks, used to evaluate performance in one of the region's most prominent scripts.
MTEB (Massive Text Embedding Benchmark): The industry-standard global benchmark used to gauge general-purpose English embedding performance across a wide array of tasks.
Factors
Evaluation factors are categorised by task type and linguistic diversity to ensure the model's "fertility" and "nuance" are captured accurately:
Linguistic Coverage: Evaluation spans across 12 languages, including complex Brahmic scripts (Burmese, Khmer, Lao, Tamil, Thai), Latin-based SEA scripts (Indonesian, Filipino, Malay, Tetum, Vietnamese), English, and Mandarin.
Task Modality:
Retrieval/Reranking: Efficiency in finding relevant documents within a large corpus.
Semantic Textual Similarity (STS): Precision in sentence-level semantic alignment.
Clustering & Classification: Ability to group or categorise text based on latent semantic meaning.
Summarisation & Bitext Mining: High-level semantic matching and cross-lingual alignment.
Architecture Efficiency: Performance is measured in the context of the ModernBERT architecture and Gemma 3 tokenizer to assess computational efficiency versus embedding quality.
Metrics
To provide a standardised view of performance, we report the following metrics across the benchmark suites:
Carbon emission was estimated using the fact sheet from TRG Datacenters.
Hardware Type: Nvidia H200 140GB GPUs
Hours used: 384 hrs
Cloud Provider: SMC H200
Compute Region: Singapore
Carbon Emitted: appx. 108.05 kg CO2 e
Technical Specifications
Model Architecture and Objective
SEA-LION-ModernBERT-Embedding is an embedding model using the ModernBERT architecture.
Parameter
SEA-LION-ModernBERT-Embedding
Layers
22 in 300M / 28 in 600M
d_model
768 in 300M / 1024 in 600M
head_dim
12 in 300M / 16 in 600M
Vocabulary
262144
Sequence Length
8k
Compute Infrastructure
Hardware
Hardware Type: Nvidia H200 140GB GPUs
Cloud Provider: SMC H200
Software
SEA-LION-ModernBERT-Embedding was trained using Sentence-Transformers.
Glossary
SEA-BED: Southeast Asia Embedding Benchmark – a comprehensive evaluation suite for embedding models on SEA languages.
Asymmetric Retrieval: Retrieval tasks where query and document formulations differ.
Mean Pooling: Aggregating token embeddings by averaging (weighted by attention mask) to produce a fixed-size sentence representation.
More Information
While this model supports masked language modeling, it is primarily optimised via contrastive fine-tuning for downstream tasks such as sequence classification, token classification, or question answering. Please note that these weights have not been specifically aligned for safety; therefore, developers should implement their own safety evaluations and security measures. The authors disclaim all liability for any claims, damages, or other liabilities arising from the use of the released code or weights.
Ahmed Dabeer, Ahn Jeongmi, Antonyrex Sajeban, Chan Hok Teng Adwin, Cheng Zi Yi Nicholas, Choa Hsueh Mei Esther, Heng Jonathan, Huang Yuli, Jann Railey Estrada Montalan, Lee Chwan Ren, Leong Wai Yi, Leong Wei Qi, Liew Rachel, Limkonchotiwat Peerat, Muhammad Ridzuan Bin Mokhtar, Nagarajan Karthik, Ng Boon Cheong Raymond, Ngee Chia Tai, Ngui Jian Gang, Nguyen Thanh Ngan, Ong Tat-Wee David, Ong Zhi Hao, Pereira Mark, Poon Joseph, Rengarajan Hamsawardhini, Siow Wei Kang Bryan, Susanto Yosephine, Sutaveephamochanon Anocha, Tan Choon Meng, Tan Chor Phin Evelyn, Tan Siao Wei Jessica, Tan Yixian, Tee Jun Yun, Teng Kok Wai Walter, Teo Eng Sipp Leslie, Tjhi William, Wu Donghang, Yeo Yeow Tong, Yong Xianbin, Zhang Haoyang, Zhang Zhou
Acknowledgement
This project is supported by the National Research Foundation Singapore and Infocomm Media Development Authority (IMDA), Singapore under its National Large Language Model Funding Initiative.
