The model is explicitly identified as being built on the MosaicML Pretrained Transformer (MPT) architecture, a decoder-only framework. Technical documentation specifies the use of Grouped Query Attention (GQA) and Rotary Positional Embeddings (RoPE). Detailed layer configurations (32 layers, d_model 4096) are publicly available in the official documentation and model cards. The training methodology, including the use of MosaicML Composer for pre-training and LoRA for instruction tuning, is well-documented.

AI Singapore provides a detailed breakdown of the 980B token pre-training dataset, including specific percentages for English (58.2%), Chinese (9.3%), and various Southeast Asian languages (e.g., Vietnamese 6.5%, Indonesian 1.5%). Sources like RefinedWeb, mC4, and The Stack are named. For the instruction-tuning phase, the documentation specifies the use of thousands of English and Indonesian pairs, though the exact public availability of the full instruction dataset is limited to descriptions of curation and native-speaker verification.

The model uses a custom SEABPETokenizer with a clearly stated vocabulary size of 256,000 tokens. The tokenizer is publicly accessible on Hugging Face, and its development (trained on 20M lines from a cleaned mC4 corpus) is documented. The use of SentencePiece and the BPE approach are verified, and the tokenizer is specifically designed to reduce tokenization overhead for Southeast Asian scripts, which is a verifiable technical claim.

The model is a dense transformer with 7.1 billion total parameters. Detailed architectural parameters are provided in official documentation (e.g., 32 layers, head_dim 32, d_model 4096). As a dense model, active parameters equal total parameters, and there is no ambiguity regarding MoE structures. The impact of the large vocabulary on the embedding layer's parameter share is also implicitly verifiable through the provided architecture specs.

The documentation provides specific hardware details: 32 AWS EC2 p4d.24xlarge instances utilizing 256 NVIDIA A100 40GB GPUs. The training duration is explicitly stated as 22 days for the 7B variant. While a specific carbon footprint calculation is not provided in the primary model card, the disclosure of hardware type, count, and duration allows for independent estimation of energy consumption.

The model was evaluated using the BHASA benchmark, which has associated public documentation and GitHub repositories. Zero-shot evaluation settings and sample sizes (100-1000 instances) are disclosed. However, while the benchmark framework is public, the specific prompt templates used for all 11 languages are not fully centralized in the model card, requiring users to reference external papers (BHASA) for full reproduction details.

The model consistently identifies as SEA-LION and is transparent about its versioning (v1) and its specific focus on Southeast Asian languages. It does not attempt to mimic other models like GPT-4 in its official documentation or system prompts. It clearly states its purpose as a regional specialist and acknowledges its limitations regarding safety alignment in the base version.

The model is released under the Apache-2.0 license, which is a standard, permissive open-source license. Documentation clearly distinguishes between the commercially permissive version (SEA-LION-7B-Instruct) and the research-only version (SEA-LION-7B-Instruct-Research, which used CC BY-NC-SA 4.0). This distinction prevents licensing ambiguity for commercial developers.

Official documentation provides VRAM estimates, noting that approximately 30GB of VRAM is required to load the model in full precision, which exceeds standard consumer hardware. While it mentions the availability of GGUF versions for quantization, it lacks a detailed matrix of VRAM requirements across different quantization levels (Q4, Q8, etc.) and their specific accuracy trade-offs.

The model uses a versioning system (v1), and AI Singapore maintains a roadmap for future versions (v2, v3). However, a detailed, granular changelog for minor weight updates or specific drift assessments between sub-versions is not prominently featured. The transition from 'Research' to 'Instruct' variants is documented, but semantic versioning for the weights themselves is less rigorous than for the software libraries.