Sahabat-AI-Gemma2-9B

The model is explicitly identified as a continued pre-training (CPT) and instruction-tuned variant of Google's Gemma 2 9B. The architecture is well-documented as a dense decoder-only transformer with 42 layers, utilizing Grouped-Query Attention (GQA), GeGLU activation, and interleaved global/local sliding window attention. The transition from the base SEA-Lionv3 model to Sahabat-AI is clearly stated in official model cards on Hugging Face and NVIDIA's documentation.

The training data volume is disclosed as approximately 50B tokens for the CPT phase and specific counts for instruction tuning (448k Indonesian, 96k Javanese, 98k Sundanese, and 129k English pairs). While the general sources (web, curated instructions, synthetic data) and collaborators (AI Singapore, Indonesian universities, media groups) are named, a precise percentage breakdown of the 50B token pre-training corpus is missing, and the raw data is not public.

The model uses the standard Gemma 2 tokenizer, which is publicly accessible and well-documented with a vocabulary size of 256,000 tokens. The documentation explicitly confirms the use of the default tokenizer to maintain compatibility with the Gemma ecosystem, and its performance across the target Indonesian dialects is verified through the SEA HELM benchmark results.

The model's parameter count is precisely stated as 9.24 billion. As a dense architecture, all parameters are active during inference. Detailed architectural specifications including the number of layers (42), hidden dimension size, and attention head configuration (16 query, 8 KV) are available through the base Gemma 2 documentation and inherited by this variant.

Compute details are provided for both the CPT and fine-tuning phases. The CPT phase utilized 32x NVIDIA H100 80GB GPUs for 7 days using MosaicML Composer. The fine-tuning and alignment phases took 4 and 2 hours respectively on 8x H100 GPUs. While hardware and duration are clear, specific energy consumption or carbon footprint calculations were not provided in the official documentation.

Evaluation results are provided for the SEA HELM (BHASA) benchmark and standard English benchmarks (IFEval, MMLU-Pro). The documentation specifies the use of native prompts and zero-shot/five-shot settings. However, the exact prompt templates and full evaluation code are not publicly hosted in a dedicated repository, and the team notes discrepancies with official leaderboards due to inference platform differences (vLLM vs. Transformers).

The model consistently identifies as part of the Sahabat-AI family and acknowledges its Gemma 2 9B foundation. It is transparent about its specialization for Indonesian regional languages. There are no documented instances of the model claiming to be a different architecture or misrepresenting its origins in official testing or documentation.

The model is released under the Gemma Community License, which is a permissive open-weights license allowing for commercial use but subject to Google's specific terms. While the license is clear, there is some minor ambiguity in third-party repositories (like Ollama) occasionally labeling it as Apache 2.0, though official sources consistently cite the Gemma Community License.

Detailed VRAM requirements are available for various configurations, including FP16 (approx. 21.34 GB) and quantized versions (e.g., Q4_K_M at 6.6 GB). Documentation provides specific GPU recommendations (RTX 3090/A6000) and notes the impact of context length (8192 tokens) on memory usage, offering high transparency for deployment.

The model uses basic versioning (v1), and the transition from 'v1.0' to 'v1' is documented in Hugging Face commit histories. However, there is no formal semantic versioning system or public changelog detailing specific weight updates or performance drift over time. The project is relatively new, so long-term tracking is not yet established.