The model is explicitly identified as a continued pre-training (CPT) variant of the Meta Llama 3 architecture. Documentation specifies it is a dense, decoder-only Transformer with 32 layers, 4096 hidden dimensions, and Grouped Query Attention (GQA). The training methodology is described as a combination of full parameter fine-tuning, on-policy alignment, and model merging using the NVIDIA NeMo framework. While the base architecture is well-documented, the specific modifications for the CPT phase are described at a high level without a full technical paper detailing the exact layer-wise changes if any.

The model card provides specific token counts for the CPT phase (approximately 50 billion tokens) and instruction tuning (448,000 Indonesian, 96,000 Javanese, 98,000 Sundanese, and 129,000 English pairs). However, the specific sources of the 50B CPT tokens are not disclosed beyond 'publicly available online data' and 'synthetic instructions'. There is no detailed breakdown of the web-to-code ratio or specific filtering/cleaning methodologies provided, which are critical for high-tier transparency.

The model utilizes the default Llama 3 tokenizer with a vocabulary size of 128,000 tokens. This tokenizer is publicly accessible and its performance on the target languages (Indonesian, Javanese, Sundanese) is verifiable through the weights. The documentation confirms the context length is 8192 tokens, though some inference engines like vLLM may cap it at 4096. The alignment between the tokenizer and the claimed language support is well-documented.

The model is clearly stated to have 8.03 billion total parameters. As a dense architecture, all parameters are active during inference. The architectural breakdown (layers, hidden dims, attention heads) is provided in the technical specifications. There is no ambiguity regarding MoE active vs. total parameters, and the parameter count is consistent across official sources.

The documentation provides specific hardware details (8x H100-80GB GPUs) and durations for the fine-tuning (4 hours) and alignment (2 hours) phases. However, the compute resources used for the 50B token continued pre-training phase are not explicitly detailed in terms of total GPU hours or hardware specifications. Carbon footprint and total estimated cost for the entire development cycle are missing.

Evaluation results are provided for the SEA HELM (BHASA) benchmark and standard English benchmarks (IFEval, MMLU-Pro). While the benchmarks are named and some methodology is described (zero-shot with native prompts), the exact evaluation code and full prompt sets are not publicly linked in the repository. The documentation notes discrepancies with official leaderboards due to inference engine differences (vLLM vs. Transformers), which adds a layer of complexity to independent verification.

The model consistently identifies itself as Sahabat-AI and is transparent about its origins as a fine-tuned Llama 3 variant. It does not claim to be a different model (like GPT-4) and provides clear versioning (v1). The model card explicitly lists its intended use cases and limitations, including a lack of safety alignment, which demonstrates high honesty regarding its identity and capabilities.

The model is released under the Llama 3.1 Community License. This license is publicly available and outlines terms for commercial use (up to 700M monthly active users) and redistribution. However, there is a slight discrepancy in documentation where some files refer to the 'Llama 3 Community License' while others mention 'Llama 3.1', which could lead to minor legal ambiguity regarding specific derivative work terms.

Basic VRAM requirements are provided through third-party and community documentation (e.g., ~16GB for FP16, ~5GB for 4-bit quantization). The official documentation mentions the use of vLLM and Hugging Face Transformers but lacks a comprehensive official table of VRAM vs. context length scaling or specific quantization accuracy trade-off data provided directly by the developers.

The model uses a versioning string (v1), but there is no public changelog or detailed version history tracking changes between internal iterations or checkpoints. There is no formal mechanism described for tracking or notifying users of behavioral drift, and previous versions of the weights are not easily accessible in a structured historical archive.