Sahabat-AI-Gemma2-9B-Instruct

The model is explicitly identified as a continued pre-training (CPT) and instruction-tuned variant of Google's Gemma 2 9B architecture. Documentation clearly details the transition from the base Gemma 2 to the CPT version (SEA-Lionv3 base) and finally to the Sahabat-AI-Gemma2-9B-Instruct. Technical refinements such as Grouped-Query Attention (GQA), RMSNorm, and logit soft-capping are well-documented. The training methodology, including full parameter fine-tuning, on-policy alignment, and model merging, is publicly disclosed in the model card and associated technical descriptions.

The model provides specific counts for its instruction-tuning datasets: 448,000 Indonesian, 96,000 Javanese, 98,000 Sundanese, and 129,000 English instruction-completion pairs. It also notes that the continued pre-training phase involved approximately 50 billion tokens. However, while categories (synthetic, hand-curated, publicly available) are mentioned, the specific raw data sources and the exact mixture of the 50B token CPT corpus are not fully itemized, preventing a higher score.

The model utilizes the standard Gemma 2 tokenizer, which has a well-documented vocabulary size of 256,128 tokens. This tokenizer is publicly accessible via the Hugging Face repository. The documentation confirms the use of this default tokenizer and its alignment with the model's architectural requirements, though specific analysis of its efficiency on regional Indonesian dialects (Javanese/Sundanese) compared to standard Indonesian is not deeply detailed in the public documentation.

The model's parameter count is clearly stated as 9.24 billion. As a dense model, all parameters are active during inference. The architectural breakdown is inherited from the well-documented Gemma 2 9B framework, including details on attention mechanisms and layer normalization. The use of '9B' in the name is accurate to its total parameter density.

The documentation provides specific hardware and duration details: fine-tuning took approximately 4 hours and alignment took 2 hours, both conducted on a cluster of 8x NVIDIA H100-80GB GPUs. While the specific compute for the 50B token continued pre-training phase is less granularly detailed than the instruction-tuning phase, the disclosure of the final tuning stages is exemplary for transparency.

Evaluation results are provided for the SEA HELM (BHASA) benchmark and IFEval, with specific scores for tasks like sentiment analysis and toxicity detection. The documentation specifies the use of zero-shot and few-shot (5-shot) prompting with native prompts. However, the exact evaluation code and the specific localized versions of the IFEval prompts are not fully public, and the model card notes discrepancies between internal vLLM-based testing and Hugging Face Leaderboard results due to context window limitations.

The model consistently identifies itself as part of the Sahabat-AI family and correctly attributes its lineage to the Gemma 2 architecture. There is no evidence of identity confusion or claims of being a competitor's model. The documentation and model card maintain a coherent identity across platforms (Hugging Face, NVIDIA NGC, and GitHub).

The model is released under the Gemma Community License, which is a standard, publicly available license allowing for both commercial and non-commercial use with specific attribution and usage restrictions. The license terms are clearly stated on the Hugging Face repository and the NVIDIA NGC catalog. There are no conflicting terms between the weights and the provided code.

VRAM requirements are well-documented across different platforms, with specific estimates for FP16 (approx. 18-21GB) and various quantization levels (e.g., Q4_0 requiring ~5.4GB). Recommended hardware (e.g., RTX 3090 for 24GB total VRAM) is provided. The documentation also notes the context length limit (8192 tokens) and its impact on memory, though detailed scaling curves for context length are not provided.

The model uses a 'v1' designation, and the release date is clearly tracked. However, there is no formal semantic versioning system or a detailed public changelog for minor updates or weight adjustments. While the repository is maintained, the infrastructure for tracking behavioral drift or accessing historical versions of the CPT weights is limited.