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-v4-27B has undergone post-training using a QA pairs dataset in Burmese, English,
Indonesian, Khmer, Lao, Malay, Tagalog, Tamil, Thai and Vietnamese, comprising approximately 10M samples in total, to create Gemma-SEA-LION-v4-27B-IT.
Gemma-SEA-LION-v4-27B-IT inherits Gemma 3's:
Large 128K context length
Image and text understanding capabilities, including document comprehension, visual Q&A, and image-grounded reasoning
Advanced function calling and structured outputs to allow for seamless integration into larger systems
Model Details
Model Description
SEA-LION stands for Southeast Asian Languages In One Network.
We performed post-training in English and SEA languages on Gemma-SEA-LION-v4-27B, a decoder model using the Gemma 3 architecture, to create Gemma-SEA-LION-v4-27B-IT.
For tokenization, the model employs the default tokenizer used in Gemma 3 27B IT.
As of 25 Aug 2025, Gemma-SEA-LION-v4-27B-IT excels at Southeast Asian (SEA) tasks when compared to other open models
with fewer than 200 billion parameters and demonstrates performance comparable to that of larger and top closed models.
For detailed rankings, please refer to the leaderboard.
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.
Limitations
In terms of vision capability, Gemma-SEA-LION-v4-27B-IT has been trained and fine-tuned exclusively on the text back-end.
As a result, its vision capabilities are expected to be comparable to those of Gemma 3 IT 27B,
and may not exhibit significant improvements or differences in this area. 🤗 google/gemma-3-27b-it
How to Get Started with the Model
Use the code below to get started with the model.
Use the code below to get started with the model using the 🤗 Transformers library.
python
1from transformers import pipeline
2import torch
34pipe = pipeline(5"text-generation",6 model="aisingapore/Gemma-SEA-LION-v4-27B-IT",7 device="cuda",8 torch_dtype=torch.bfloat16
9)1011messages =[12{13"role":"system",14"content":[{"type":"text","text":"You are a helpful assistant."}]15},16{17"role":"user",18"content":[19{"type":"text","text":"Write a poem on southeast asian countries in Indonesian."}20]21}22]2324output = pipe(text=messages, max_new_tokens=200)25print(output[0]["generated_text"][-1]["content"])
Training Details
Training Datasets:
The instruction fine-tuning dataset combines our SEA-Instruct, Infinity-Instruct,
and OpenMath-Instruct 2 with open-source datasets. For the Online RL datasets, open sourced datasets such as
nvidia/Llama-Nemotron-Post-Training-Dataset (RL set) and zwhe99/DeepMath-103K were used. For alignment, rejected-chosen pairs are generated
from the target model, with the chosen responses obtained by rewriting and improving upon
the rejected outputs.
Prompt sampling is guided by a gradient-based analysis process.
Training Procedure
Training Hyperparameters
Training regime:
Our post-training workflow consists of multiple stages: instruction fine-tuning,
model merging, online RL for both instruction following and math using DRGPPO,
and then followed by on-policy alignment via APO.
Evaluation
Testing Data, Factors & Metrics
Testing Data
We evaluated Gemma-SEA-LION-v4-27B-IT on general language, multi-turn chat and instruction-following capabilities.
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.
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:
Task
Metric
Sentiment Analysis
Accuracy
Extractive QA (ID, VI, TH, TA)
ChrF++
MCQ-QA (TL, MY, MS)
Accuracy
Metaphor
Accuracy
Abstractive Summarisation
Rouge-L
Translations
MetricX-24 score (with reference)
Causal Reasoning
Accuracy
Natural Language Inference
Accuracy
LINDSEA
Accuracy
Global MMLU Lite
Accuracy
Kalahi
Accuracy
SEA-IFEval
Accuracy
SEA-MTBench
Win rate against a reference
Toxicity Detection
Accuracy
Results
Leaderboard results on SEA-HELM snapedshot on 25 Aug 2025!
For details on Gemma-SEA-LION-v4-27B-IT performance, please refer to the SEA-HELM leaderboard, Leaderboard results on SEA-HELM.
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.
AI Singapore is a national programme supported by the National Research Foundation, Singapore and hosted by the National University of Singapore.
Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of the National Research Foundation or the National University of Singapore.
