SEA-LION is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for the Southeast Asia (SEA) region.
Apertus-SEA-LION-v4-8B-IT is a 8-billion parameter model built upon the Apertus-8B-Instruct architecture. To ensure domain adaptation for the region, the model underwent rigorous post-training on a curated dataset of approximately 6.4 million instruction-text pairs.
This extensive post-training instills multilingual and multicultural fluency, covering key SEA languages such as Indonesian, Vietnamese, Thai, Filipino, Tamil, Burmese, Malay. This curated dataset also includes a filtered open sourced set of tool-calling instruction-text pairs to impart these capabilities, in addition to linguistic fluency.
Apertus-SEA-LION-v4-8B-IT is designed as a fully open model to align with this core philosophy, we have released the datasets used for post-training, as well as the evaluation codes and datasets used to evaluate the model.
These resources can be accessed via the link below.
SEA-LION stands for Southeast Asian Languages In One Network.
We performed Post-Training in English and SEA languages on Apertus-8B-Instruct-2509, a decoder model using the Apertus architecture, to create Apertus-SEA-LION-v4-8B-IT.
For tokenization, the model employs the default tokenizer used in Apertus-8B-Instruct-2509.
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.
How to Get Started with the Model
Use the code below to get started with the model with 🤗 Transformers libraries.
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "aisingapore/Apertus-SEA-LION-v4-8B-IT"
model = AutoModelForCausalLM.from_pretrained(
model_id, device_map="auto"
).eval()
tokenizer = AutoTokenizer.from_pretrained(model_id)
messages = [
{
"role": "user",
"content": "Any flats for sale in Tampines?"
}
]
tools = [
{
"type": "function",
"function": {
"name": "search_hdb_listings",
"description": "Search for HDB flats available for sale",
"parameters": {
"type": "object",
"properties": {
"location": {"type": "string", "description": "Town or area name"},
"flat_type": {"type": "string", "description": "Flat type e.g. 3-room, 4-room, 5-room"},
"max_price": {"type": "number", "description": "Maximum price in SGD"}
},
"required": ["location", "flat_type"]
}
}
},
{
"type": "function",
"function": {
"name": "calculate_mortgage",
"description": "Calculate estimated monthly mortgage payment",
"parameters": {
"type": "object",
"properties": {
"loan_amount": {"type": "number", "description": "Loan amount in SGD"},
"interest_rate": {"type": "number", "description": "Annual interest rate as percentage"},
"loan_tenure_years": {"type": "integer", "description": "Loan period in years"}
},
"required": ["loan_amount"]
}
}
}
]
inputs = tokenizer.apply_chat_template(
messages, tools=tools, add_generation_prompt=True, tokenize=True,
return_dict=True, return_tensors="pt"
).to(model.device)
input_len = inputs["input_ids"].shape[-1]
with torch.inference_mode():
generation = model.generate(**inputs, max_new_tokens=512, do_sample=False)
generation = generation[0][input_len:]
decoded = tokenizer.decode(generation, skip_special_tokens=True)
print(decoded)
Output:
# I'll search for HDB flats available for sale in Tampines. Let me check the listings for you.
# <tool_call>
# {"name": "search_hdb_listings", "arguments": {"location": "Tampines", "flat_type": "3-room", "max_price": 500000}}
# </tool_call>
Training Datasets:
The instruction fine-tuning text dataset comprises of a collection of OSS & synthetic data. The datasets used for Post Training can be accessed via the link below.
Training Regime: Our Post Training workflow consists of Instruction Fine Tuning and Distillation (from Apertus-70B-Instruct-2509).
Evaluation
Testing Data, Factors & Metrics
We evaluated Apertus-SEA-LION-v4-8B-IT on general language capabilities and LLM-specific capabilities using SEA-HELM.
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
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
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-4.1-2025-04-14 as the judge model and compare against gpt-4.1-2025-04-14 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:
For details on Apertus-SEA-LION-v4-8B-IT performance, please refer to the SEA-HELM leaderboard, https://leaderboard.sea-lion.ai/ .
Tool calling
We evaluated the tool calling capabilities of our model using the Berkeley Function Calling Leaderboard (BFCL) V4 evaluation.
Factors
The evaluation was done using the codes from the BFCL v4 repository. Modifications were made to the agentic web search task:
The Brave Search API was used instead of the DuckDuckGo Search API. The Brave Search API is also a privacy focused search engine and is similar to the DuckDuckGo Search API.
Note: Apertus-8b was run in prompt mode as it does not support function calling.
Results
ToolcallingResults!
Technical Specifications
Software Environment & Requirements
Library
Version
Transformers
4.57.1
PyTorch
2.8.0 + cu129
Datasets
4.4.2
Tokenizers
0.22.1
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 Zhou, Imanol Schlag (Swiss AI), Antoine Bosselut (Swiss AI) and Martin Jaggi (Swiss AI)
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.
