Model Card for SEA-LION-ModernBERT-600M-checkpoints

Last update: 2026-03-16

SEA-LION is a collection of Large Language Models (LLMs) which have been pretrained and fine-tuned for the Southeast Asia (SEA) region.

This encoder-only model leverages the advanced ModernBERT architecture combined with the Gemma 3 SentencePiece tokenizer. The adoption of the Gemma 3 tokenizer on top of ModernBERT allows the model to achieve highly efficient and culturally nuanced text processing. This combination significantly improves the tokenization fertility and compression rates for complex regional scripts and diverse Southeast Asian languages, enabling the model to handle longer context windows and cross-lingual tasks with greater computational efficiency.

To achieve this level of performance, the model was developed through a rigorous, multi-stage training pipeline. The foundation was established through extensive pre-training on 2 Trillion (2T) tokens, followed by a mid-training phase on an additional 1 Trillion (1T) tokens. Both of these massive training phases comprehensively covered code alongside 13 specific languages: Burmese, Chinese, English, Filipino, Indonesian, Javanese, Khmer, Lao, Malay, Sundanese, Tamil, Thai, and Vietnamese.

Model Details

Model Description

The SEA-LION-ModernBERT-600M-checkpoints model is built on the ModernBERT-large architecture and has a vocabulary size of 262K.

For tokenization, the model employs our custom Gemma3 tokenizer, which has excellent performance for SEA languages, ensuring optimal model performance.

• Developed by: AI Products Pillar, AI Singapore
• Funded by: Singapore NRF
• Shared by: AI Products Pillar, AI Singapore
• Model type: Encoder
• Context length: 8k
• Languages: Burmese, Chinese, English, Filipino, Indonesian, Javanese, Khmer, Lao, Malay, Sundanese, Tamil, Thai, and Vietnamese
• License: MIT

Model Sources

• Repository: The weights for this model and its various training stages are being released to support transparency, research, and diverse downstream applications. Link to HF Repo

Uses

This model card details one of the variants available within this collection.

Model Variant	Model Repository	Suggesting Applications & Use Cases
Fine-tuned Embedding Models	- aisingapore/SEA-LION-E5-Embedding-600M - aisingapore/SEA-LION-ModernBERT-Embedding-300M - aisingapore/SEA-LION-ModernBERT-Embedding-600M	- Retrieval-Augmented Generation (RAG) - Information retrieval, and search - Similarity comparisons
Pre-trained Encoder Models	- aisingapore/SEA-LION-ModernBERT-300M - aisingapore/SEA-LION-ModernBERT-600M	- Fill mask - Text classification - Fine-tuning for downstream tasks (e.g., sentiment analysis, classification).
Pre-trained Model Checkpoints	- aisingapore/SEA-LION-ModernBERT-300M-checkpoints - aisingapore/SEA-LION-ModernBERT-600M-checkpoints	- Continued Pre-Training (CPT) - Fine-tuning for downstream tasks (e.g., sentiment analysis, classification).

This repository contains model variants available of the 600M checkpoints.

Model Variant	Suggesting Applications & Use Cases
stage1-pre-training/SEA-LION-PT-600M.pt	Composer checkpoint from the Pre-Training Stage suitable for continued pre-training (CPT).
stage1-pre-training/SEA-LION-PT-600M	Folder for the HuggingFace checkpoints from the Pre-Training Stage, suitable for continued pre-training or fine tuning.
stage2-mid-training/SEA-LION-MT-600M-w-decay.pt	Composer checkpoint from the Mid-Training stage with learning rate annealing suitable for fine tuning with learning rate warmup.
stage2-mid-training/SEA-LION-MT-600M-wo-decay.pt	Composer checkpoint from the Mid-Training stage without learning rate annealing suitable for continued pre-training (CPT) and fine tuning without learning rate warmup.
stage2-mid-training/SEA-LION-MT-600M-wo-decay	Folder for the HuggingFace checkpoints from the Mid-Training stage without learning rate annealing. suitable for continued pre-training (CPT) and fine tuning without learning rate warmup.

Note:
1. For stage2-mid-training checkpoints with learning rate annealing, please refer to aisingapore/SEA-LION-ModernBERT-600M.
2. If you are deploying our models for your specific use case, we would love to hear from you! Please feel free to contact us to share your experience or explore potential collaborations.

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.

How to Get Started with the Model

Use the code below to download the model locally.

pip install -U transformers>=4.48.0

#########################
## Download checkpoints locally for continued pre-training or fine tuning
#########################
from huggingface_hub import snapshot_download

# Download the stage-1-pre-training Huggingface checkpoint
snapshot_download(
  "aisingapore/SEA-LION-ModernBERT-600M-checkpoints",
  repo_type="model",
  allow_patterns=["stage1-pre-training/SEA-LION-PT-600M/*"],
  local_dir="checkpoints"
)

# Download the stage-1-pre-training Composer checkpoint
snapshot_download(
  "aisingapore/SEA-LION-ModernBERT-600M-checkpoints",
  repo_type="model",
  allow_patterns=["stage1-pre-training/SEA-LION-PT-600M.pt"],
  local_dir="checkpoints"
)

import torch
from transformers import pipeline

pipeline = pipeline(
    task="fill-mask",
    model="checkpoints/stage1-pre-training/SEA-LION-PT-600M",
    dtype=torch.float16,
    device=0
)
pipeline("Plants create  through a process known as photosynthesis.")

Note: To get started with Continued Pre-Training of the Composer checkpoints, we recommend refering to this guide.

---

Training Details

The models are pre-training from scratch through a two-phase pipeline, beginning with an extensive initial stage on 2 trillion tokens, followed by a mid-training phase on an additional 1 trillion tokens. Both phases incorporated a diverse dataset covering programming code and 13 languages: Burmese, Chinese, English, Filipino, Indonesian, Javanese, Khmer, Lao, Malay, Sundanese, Tamil, Thai, and Vietnamese.

Training Data

SEA-LION-ModernBERT-600M-checkpoints was pre-trained from scratch on a number of trillion token corpus with the following linguistic and thematic distribution:

Data Source	Percentage
code	10%
EN - English	35%
ID - Indonesian	8%
JV - Javanese	0.5%
KM - Khmer	1.5%
LO - Lao	0.5%
MS - Malay	4.75%
MY - Burmese	1.75%
SU - Sundanese	0.5%
TA - Tamil	4.5%
TH - Thai	8%
TL - Filipino	2.5%
VI - Vietnamese	8.5%
ZH - Chinese	14%

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 10+ languages, including complex Brahmic scripts (Burmese, Khmer, Lao, Tamil, Thai) and Latin-based SEA scripts (Indonesian, Filipino, Malay, Tetum, Vietnamese).

• 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 categorize 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 standardized view of performance, we report the following metrics across the benchmark suites:

• Classification: F1-score.
• Multi-label Classification: F1-score
• Pair Classification: Average Precision (AP).
• Semantic Textual Similarity (STS): Cosine similarity scores.
• Clustering: V-Measure Score.
• Bitext Mining: F1-score.
• Retrieval & Reranking: NDCG@10 (Primary) and MAP.
• Instruction Retrieval: NNDCG@5

Results

Performance comparison of embedding models on SEA-BED (https://leaderboard.sea-lion.ai/embedding/SEA). Captured on 13/03/2026 02:50pm.

Environmental Impact

Carbon emission was estimated using the fact sheet from TRG Datacenters.

• Hardware Type: Nvidia H200 140GB GPUs
• Hours used: 3,680 hrs
• Cloud Provider: SMC H200
• Compute Region: Singapore
• Carbon Emitted: appx. 1,035.552 kg CO2 e

Technical Specifications

Model Architecture and Objective

SEA-LION-cptr-600M is an encoder model using the ModernBERT architecture.

Parameter	SEA-LION-ModernBERT
Layers	28
d_model	1024
head_dim	16
Vocabulary	262144
Sequence Length	8k

Compute Infrastructure

Hardware

• Hardware Type: Nvidia H200 140GB GPUs
• Cloud Provider: SMC H200

Software

SEA-LION was trained using the ModernBERT code base which is powered by the Composer training framework from MosaicML.

More Information

This is the repository for the commercial fine-tuned model. The model has not been aligned for safety. Developers and users should perform their own safety fine-tuning and related security measures. In no event shall the authors be held liable for any claims, damages, or other liabilities arising from the use of the released weights and codes.

For more info, please contact us at sealion@aisingapore.org

Team

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.

Contact

sealion@aisingapore.org