gemma2-9b-cpt-sahabatai-v1-base

Gemma2 9B CPT Sahabat-AI v1

Sahabat-AI (Indonesian language for “close friends”) is a collection of Large Language Models (LLMs) which has been pretrained and instruct-tuned for Indonesian language and its various dialects. Sahabat-AI ecosystem is co-initiated by Indonesian tech and telecommunication companies: GoTo Group and Indosat Ooredoo Hutchison.

This is the card for the Gemma2 9B CPT Sahabat-AI v1 base model which has undergone continued pre-training from the Gemma2 9B CPT SEA-Lionv3 base model.

Model Details

Model Description

The continued pre-training data for Gemma2 9B CPT Sahabat-AI v1 base model encompasses approximately 50B tokens.

• Co-initiated by: PT GoTo Gojek Tokopedia Tbk, Indosat Ooredoo Hutchison
• Developed by: PT GoTo Gojek Tokopedia Tbk, AI Singapore
• Model type: Decoder
• Languages: English, Indonesian, Javanese, Sundanese
• License: Gemma Community License

For tokenisation, the model employs the default tokenizer used in Gemma-2-9B. The model has a context length of 8192.

Benchmark Performance

We evaluated Gemma2 9B CPT Sahabat-AI v1 base model on general language capabilities.

General Language Capabilities

For the evaluation of general language capabilities, we employed the

• SEA HELM (also known as BHASA) evaluation benchmark across a variety of tasks.
  • These tasks include Question Answering (QA), Sentiment Analysis (Sentiment), Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng), Abstractive Summarization (Summ), Causal Reasoning (Causal) and Natural Language Inference (NLI).
  • We also added support for Javanese and Sundanese for the BHASA tasks whenever applicable
• and the common English tasks from the HuggingFace LLM Leaderboard.
  • These tasks consist of IFEval, BBH, Math Lvl 5, GPQA, MuSR, and MMLU-PRO.
  • Caveat: Our results differ from the HuggingFace LLM Leaderboard because we have used VLLM as our inference platform. VLLM caps the context size at 4096 tokens while HuggingFace was set to 8192 tokens.

Note: SEA HELM is implemented using prompts to elicit answers in a strict format. For all tasks, the model is expected to provide an answer tag from which the answer is automatically extracted. For tasks where options are provided, the answer should comprise one of the pre-defined options. The scores for each task is normalised to account for baseline performance due to random chance.

The evaluation was done five-shot with native prompts on a sample of 100-1000 instances for each dataset.

Results

SEA HELM (also known as BHASA)

<table style="border-collapse: collapse; width: 100%; font-size: 10px"> <tr> <th style="border: 2px solid black; padding: 8px; font-weight: bold;">Language / Model Name [Base]</th> <th style="border: 1px solid gray; padding: 8px;">Qwen2-7B</th> <th style="border: 1px solid gray; padding: 8px;">Qwen2.5-7B</th> <th style="border: 1px solid gray; padding: 8px;">Llama-3-8B</th> <th style="border: 1px solid gray; padding: 8px;">Llama-3.1-8B</th> <th style="border: 1px solid gray; padding: 8px;">sea-lionv2.1-8B</th> <th style="border: 1px solid gray; padding: 8px;">gemma-2-9B</th> <th style="border: 1px solid gray; padding: 8px;">sea-lionv3-9B</th> <th style="border: 1px solid gray; padding: 8px;">sahabatai-v1-8B</th> <th style="border: 2px solid black; padding: 8px;">sahabatai-v1-9B</th> </tr> <tr> <td style="border: 2px solid black; padding: 8px; font-weight: bold;">Overall (Bahasa Indonesia + Javanese + Sundanese)</td> <td style="border: 1px solid gray; padding: 8px;">42.776</td> <td style="border: 1px solid gray; padding: 8px;">46.245</td> <td style="border: 1px solid gray; padding: 8px;">49.160</td> <td style="border: 1px solid gray; padding: 8px;">49.577</td> <td style="border: 1px solid gray; padding: 8px;">48.602</td> <td style="border: 1px solid gray; padding: 8px;">58.972</td> <td style="border: 1px solid gray; padding: 8px;">60.913</td> <td style="border: 1px solid gray; padding: 8px;">59.437</td> <td style="border: 2px solid black; padding: 8px; background-color: lightgreen;">64.123</td> </tr> <tr> <td style="border: 2px solid black; padding: 8px; font-weight: bold;">Bahasa Indonesia</td> <td style="border: 1px solid gray; padding: 8px;">49.341</td> <td style="border: 1px solid gray; padding: 8px;">55.913</td> <td style="border: 1px solid gray; padding: 8px;">47.865</td> <td style="border: 1px solid gray; padding: 8px;">48.110</td> <td style="border: 1px solid gray; padding: 8px;">49.154</td> <td style="border: 1px solid gray; padding: 8px;">58.572</td> <td style="border: 1px solid gray; padding: 8px; background-color: lightgreen;">62.437</td> <td style="border: 1px solid gray; padding: 8px;">53.454</td> <td style="border: 2px solid black; padding: 8px;">60.040</td> </tr> <tr> <td style="border: 2px solid black; padding: 8px; font-weight: bold;">Javanese</td> <td style="border: 1px solid gray; padding: 8px;">42.774</td> <td style="border: 1px solid gray; padding: 8px;">45.917</td> <td style="border: 1px solid gray; padding: 8px;">54.627</td> <td style="border: 1px solid gray; padding: 8px;">55.215</td> <td style="border: 1px solid gray; padding: 8px;">52.728</td> <td style="border: 1px solid gray; padding: 8px;">63.760</td> <td style="border: 1px solid gray; padding: 8px;">63.363</td> <td style="border: 1px solid gray; padding: 8px;">65.048</td> <td style="border: 2px solid black; padding: 8px; background-color: lightgreen;">69.882</td> </tr> <tr> <td style="border: 2px solid black; padding: 8px; font-weight: bold;">Sundanese</td> <td style="border: 1px solid gray; padding: 8px;">36.213</td> <td style="border: 1px solid gray; padding: 8px;">36.905</td> <td style="border: 1px solid gray; padding: 8px;">44.988</td> <td style="border: 1px solid gray; padding: 8px;">45.407</td> <td style="border: 1px solid gray; padding: 8px;">43.925</td> <td style="border: 1px solid gray; padding: 8px;">54.583</td> <td style="border: 1px solid gray; padding: 8px;">56.939</td> <td style="border: 1px solid gray; padding: 8px;">59.809</td> <td style="border: 2px solid black; padding: 8px; background-color: lightgreen;">62.446</td> </tr> </table>

English Results

<table style="border-collapse: collapse; width: 100%; font-size: 10px"> <tr> <th style="border: 1px solid gray; padding: 8px;">Model Name [BASE]</th> <th style="border: 1px solid gray; padding: 8px;">Qwen2-7B</th> <th style="border: 1px solid gray; padding: 8px;">Qwen2.5-7B</th> <th style="border: 1px solid gray; padding: 8px;">Llama-3-8B</th> <th style="border: 1px solid gray; padding: 8px;">Llama-3.1-8B</th> <th style="border: 1px solid gray; padding: 8px;">sea-lionv2.1-8B</th> <th style="border: 1px solid gray; padding: 8px;">gemma-2-9B</th> <th style="border: 1px solid gray; padding: 8px;">sea-lionv3-9B</th> <th style="border: 1px solid gray; padding: 8px;">sahabatai-v1-8B</th> <th style="border: 2px solid black; padding: 8px;">sahabatai-v1-9B</th> </tr> <tr> <td style="border: 1px solid gray; padding: 8px; font-weight: bold;">Average</td> <td style="border: 1px solid gray; padding: 8px;">23.68</td> <td style="border: 1px solid gray; padding: 8px; background-color: lightgreen;">24.65</td> <td style="border: 1px solid gray; padding: 8px;">13.56</td> <td style="border: 1px solid gray; padding: 8px;">13.69</td> <td style="border: 1px solid gray; padding: 8px;">12.77</td> <td style="border: 1px solid gray; padding: 8px;">13.34</td> <td style="border: 1px solid gray; padding: 8px;">21.99</td> <td style="border: 1px solid gray; padding: 8px;">13.92</td> <td style="border: 2px solid black; padding: 8px;">19.62</td> </tr> </table>

Training Details

Data

Gemma2 9B CPT Sahabat-AI v1 base model was continued pre-trained on 50B tokens of the following data:

Note:

• All token counts are counted using Gemma2 tokenizer
• Wiki* sources includes Wikipedia, Wiki Books, Wiki Source, Wiki Voyage and Fandom Wiki
• News* sources includes VOA, Global Voices

Infrastructure

Gemma2 9B CPT Sahabat-AI v1 was trained using MosaicML Composer on the following hardware:

Configuration

Call for Collaboration

Sahabat-AI (Indonesian language for “close friends”) a local open source Large Language Model (LLM) ecosystem in Indonesian language, co-initiated by Indonesian tech and telecommunication companies: GoTo Group and Indosat Ooredoo Hutchison. Sahabat-AI ecosystem aims to empower Indonesians who want to develop AI-based services and applications using Bahasa Indonesia and its various local dialects.

We are supported by research centers and global tech experts such as AI Singapore and Tech Mahendra to train the model to gain general language understanding.

We also collaborate with key top Indonesia universities such as University of Indonesia, Gadjah Mada University, Bogor Institute of Agriculture, Bandung Institute of Technology, including top Indonesia media groups, such as Kompas Gramedia Group and Republika to train and enrich the model in Bahasa Indonesia, ensuring optimum provision of local context and cultural relevance.

We would like to invite researchers, developers, and language enthusiasts to actively contribute to the enhancement and expansion of Sahabat-AI. Your collaborations can involve:

• Identifying and reporting technical issues
• Sharing pre-training, instruction, and preference data
• Improving documentation usability
• Proposing and implementing new model evaluation tasks and metrics

Join us in shaping the future of Sahabat-AI by sharing your expertise and insights to make these models more accessible, accurate, and versatile.

You can contribute your ideas through this form.

The Development Team (in ascending alphabetical order)

AI Singapore

Chan Adwin<br> Cheng Nicholas<br> Choa Esther<br> Huang Yuli<br> Lau Wayne<br> Lee Chwan Ren<br> Leong Wai Yi<br> Leong Wei Qi<br> Limkonchotiwat Peerat<br> Liu Bing Jie Darius<br> Montalan Jann Railey<br> Ng Boon Cheong Raymond<br> Ngui Jian Gang<br> Nguyen Thanh Ngan<br> Ong Brandon<br> Ong Tat-Wee David<br> Ong Zhi Hao<br> Rengarajan Hamsawardhini<br> Siow Bryan<br> Susanto Yosephine<br> Tai Ngee Chia<br> Tan Choon Meng<br> Teng Walter<br> Teo Eng Sipp Leslie<br> Teo Wei Yi<br> Tjhi William<br> Yeo Yeow Tong<br> Yong Xianbin<br>

PT GoTo Gojek Tokopedia Tbk

Anissa Dininta<br> Chau Shiau Ching<br> Choiri Hendra Hadhil<br> Goel Priyank<br> Saini Ajay Kumar<br> Shalev Ofir<br> Tan Daryl<br> Tep Kilian Rithi<br> Tiwari Anupam<br> Widjojo Daniel<br>

Acknowledgements

AI Singapore is a national programme supported by the National Research Foundation, Singapore and hosted by the National University of Singapore.

Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore.

Contact

For more info, please contact us using this Sahabat-AI Inquiry Form.

Disclaimer

This is the repository for the base model. The model has not been aligned for safety. Developers and users should perform their own safety fine-tuning and related security measures. In no event shall the authors be held liable for any claim, damages, or other liability arising from the use of the released weights and codes.