SEA-LION-7B-Instruct
Parameters
7.1B
Context Length
2K
Modality
Text
Architecture
Dense
License
Apache-2.0
Release Date
1 Feb 2024
Knowledge Cutoff
Sep 2023
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
1x RTX 4090
24GB VRAM
Datacenter
1x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
2,048 tokens
Consumer
1x RTX 4090
24GB VRAM
Datacenter
1x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
No evaluation benchmarks for SEA-LION-7B-Instruct available.
Rankings
Overall Rank
-
Coding Rank
-
About SEA-LION-7B-Instruct
SEA-LION-7B-Instruct is a specialized large language model designed specifically for the Southeast Asian (SEA) region, providing optimized instruction-following capabilities across a diverse range of regional languages. Developed by AI Singapore, this model is built upon the MosaicML Pretrained Transformer (MPT) architecture, a decoder-only framework engineered for efficient training and inference. The model utilizes a custom-designed SEABPETokenizer with a significant vocabulary size of 256,000, which is specifically tailored to handle the unique linguistic structures and character sets of Southeast Asian languages, thereby reducing tokenization overhead and improving semantic representation compared to generic tokenizers.
Technically, the architecture is a dense transformer that incorporates key optimizations such as Grouped Query Attention (GQA) for improved memory efficiency and performance during inference. It employs Rotary Positional Embeddings (RoPE) to facilitate better handling of long-range dependencies within its context window. The instruction-tuning phase involved training on a rigorously curated dataset of English and Indonesian instruction-completion pairs, along with smaller sets for other ASEAN languages like Malay, Thai, Vietnamese, Filipino, Tamil, Burmese, Khmer, and Lao. This tuning process was performed using parameter-efficient fine-tuning (PEFT) techniques, specifically Low-Rank Adaptation (LoRA), ensuring the model maintains its foundational knowledge while specializing in task-oriented responses.
Performance characteristics of the model center on its ability to perform natural language understanding (NLU), generation (NLG), and reasoning (NLR) tasks within a Southeast Asian cultural and linguistic context. It is particularly effective for use cases such as regional question-answering, localized sentiment analysis, and translation between English and SEA languages. By prioritizing commercially permissive and high-quality training data, the model serves as a reliable foundation for developers building AI applications that require cultural nuance and linguistic accuracy for the Singaporean and broader ASEAN markets.
Technical Specifications
Attention
Attention Structure
Multi-Head Attention
Attention Heads
32
Key-Value Heads
32
Attention Head Dimension
128
Position Embedding
Absolute Position Embedding
RoPE Theta
-
Sliding Window Attention
No
Sliding Window Size
-
Sliding Window Ratio
-
Linear Attention
-
Linear Attention Ratio
-
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Hidden Dimension Size
4,096
Number of Layers
32
FFN Intermediate Size (Dense)
16,384
Multi-Token Prediction Heads
-
Tokenizer
Vocabulary Size
256,000
About SEA-LION
Southeast Asian Languages In One Network (SEA-LION) is a family of language models developed by AI Singapore for Southeast Asian languages. The models support English, Indonesian, Malay, Thai, Vietnamese, Tagalog, Burmese, Khmer, Lao, Tamil, and Chinese. It focuses on regional linguistic patterns and is available in base and instruction-tuned variants.