SEA-LION is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for the Southeast Asia (SEA) region.
Gemma-SEA-LION-v4.5-E2B-IT built upon the gemma-4-E2B-it architecture with 2.3B effective (5.1B with embeddings). To ensure deep domain adaptation, the model underwent distillation from google/gemma-4-31B-it on an updated aisingapore/SEA-Instruct-2602, instilling multilingual and multicultural fluency across English and key SEA languages: Burmese, Indonesian, Filipino (Tagalog), Malay, Tamil, Thai, and Vietnamese.
Gemma-SEA-LION-v4.5-E2B-IT inherits the following features from Gemma 4:
Native System Prompt Support – Gemma 4 introduces native support for the system role, enabling more structured and controllable conversations.
Reasoning – Highly capable reasoning model, with configurable thinking modes.
Extended Multimodalities – Processes Text, Image with variable aspect ratio and resolution support (all models), Video, and Audio (featured natively).
Optimized for On-Device – designed for efficient local execution on laptops and mobile devices.
SEA-LION stands for Southeast Asian Languages In One Network.
We performed post-training in English and SEA languages on gemma-4-E2B-it, a multimodal learning model using the Gemma 4 architecture, to create Gemma-SEA-LION-v4.5-E2B-IT.
For tokenization, the model employs the default tokenizer used in gemma-4-E2B-it.
Developed by: AI Products Pillar, AI Singapore
Funded by: Singapore NRF
Shared by: AI Products Pillar, AI Singapore
Model type: Transformer Decoder with Vision and Audio Encoder
Training Stage: Post-Training (Logit Distillation & Model Merging)
Context length: 128k
Language(s): fine-tuned on Burmese, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese
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.
Bias, Risks, and Limitations
The model was not tested for robustness against adversarial prompting. It is important for users to be aware that our model exhibits certain limitations that warrant consideration. Like many LLMs, the model can hallucinate and occasionally generates irrelevant content, introducing fictional elements that are not grounded in the provided context. Users should also exercise caution in interpreting and validating the model's responses due to the potential inconsistencies.
Lineage & Versioning
This model card serves as an immutable record of the final released checkpoint. Consequently, the hardware specifications, compute hours, and precise dataset volumes (such as final filtered instruction counts) reported here reflect the exact production run used to generate this specific artifact. These figures may differ from the aggregate totals, pre-filtered data pools, or preliminary experimental runs (e.g., initial H100 benchmarks) documented in our accompanying research papers.
The training counts and dataset mixture details reported in this model card reflect the exact constructed training pool consumed during this specific model run. Figures may differ slightly from public dataset releases, which represent downloadable open-source subsets of the broader corpus.
How to Get Started with the Model
Use the code below to get started with the model with 🤗 Transformers libraries.
pip install -U transformers torch accelerate
from transformers import AutoProcessor, AutoModelForCausalLM
MODEL_ID = "aisingapore/Gemma-SEA-LION-v4.5-E2B-IT"
# Load model
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
dtype="auto",
device_map="auto"
)
# Prompt
messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Write a short joke about HDB flat in Singapore."},
]
# Process input
text = processor.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=False
)
inputs = processor(text=text, return_tensors="pt").to(model.device)
input_len = inputs["input_ids"].shape[-1]
# Generate output
outputs = model.generate(**inputs, max_new_tokens=1024)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
# Parse output
processor.parse_response(response)
Training Details
The instruction fine-tuning text dataset comprises a collection of OSS & synthetic data, including approximately 8.54 million instruction-text pairs and advanced tool-calling instruction pairs specifically curated for the region.
To maintain high efficiency, our post-training pipeline is centred entirely on knowledge distillation.
Evaluation
Testing Data, Factors & Metrics
We evaluated Gemma-SEA-LION-v4.5-E2B-IT on general language, multi-turn chat, instruction-following capabilities, and vision-language benchmarks.
Testing Data
General language capabilities
For the evaluation of general language capabilities, we employed the SEA-HELM evaluation benchmark across a variety of tasks. These tasks include Question Answering (QA), Sentiment Analysis (Sentiment), Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng), Abstractive Summarisation (Abssum), Causal Reasoning (Causal), Natural Language Inference (NLI), Linguistic Diagnostics (LINDSEA), Cultural Knowledge (Kalahi) and Global MMLU Lite/Thai Exam.
Instruction-following and Multi-turn Chat
We evaluated the models on Instruction-following and Multi-turn Chat capabilities with SEA-IFEval (based on IFEval) and SEA-MTBench (based on MT-Bench) respectively. The two datasets were originally in English, the linguists and native speakers in the team worked together to filter, localise and translate the datasets into the respective target languages to ensure that the examples remained reasonable, meaningful and natural.
Factors
All evaluations were run with the model specific generation parameters defined in the model config. Each evaluation comprised of 8 runs with different seeds and the final results were averaged across these runs.
For all tasks, the model was expected to provide an answer tag from which the answer was automatically extracted. For tasks where options were provided, the answer should comprise one of the pre-defined options.
The evaluation was done zero-shot with native prompts on a sample of 100-1000 instances for each dataset.
SEA-IFEval: Evaluates a model's ability to adhere to constraints provided in the prompt, for example beginning a response with a specific word/phrase or answering with a certain number of sections. Additionally, accuracy is normalised by the proportion of responses in the correct language (if the model performs the task correctly but responds in the wrong language, it is judged to have failed the task).
SEA-MTBench: Evaluates a model's ability to engage in Multi-turn (2 turns) conversations and respond in ways that align with human needs. We use gpt-oss-120b as the judge model and compare against gpt-oss-120b as the baseline model. The metric used is the weighted win rate against the baseline model (i.e. average win rate across each category: Math, Reasoning, STEM, Humanities, Roleplay, Writing, Extraction).
Metrics
The following metrics were used for text capabilities:
*We are constantly updating the leaderboard - more to come very soon!
Technical Specifications
Model Architecture
The architecture is based on the dense model architecture of Gemma4 E2B, can be found in the gemma-4-E2B-it Model Card.
More Information
This is the repository for the commercial instruction-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.
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.
