Co-initiated by: PT GoTo Gojek Tokopedia Tbk, Indosat Ooredoo Hutchison
Developed by: PT GoTo Gojek Tokopedia Tbk, AI Singapore
Model type: Decoder
Languages supported: English, Indonesian, Javanese, Sundanese, Batak Toba, Balinese
License: Llama 3.1 Community License

Model Details

Model Description

For tokenisation, the model employs the default tokenizer used in Llama 3.1 70B Instruct. The model has a context length of 128k.

Benchmark Performance

We evaluated Llama-Sahabat-AI-v2-70B-IT on both general language capabilities and instruction-following capabilities.

IndoMMLU

For the evaluation of capabilities rooted in the local Indonesian context, we employed the IndoMMLU evaluation benchmark across a variety of tasks.

These tasks include examination questions on Humanities, Indonesian language, Local languages and cultures, Social science and STEM across primary, middle, and high school levels.

Note: IndoMMLU 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.

General Language Capabilities

For the evaluation of general language capabilities, we employed the SEA-HELM 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 Summarisation (Abssum), Causal Reasoning (Causal), Natural Language Inference (NLI), and linguistic diagnostics (LINDSEA).

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 zero-shot with native prompts on a sample of 100-1000 instances for each dataset.

Instruction-following Capabilities

Since Llama-Sahabat-AI-v2-70B-IT is an instruction-following model, we also evaluated it on instruction-following capabilities with two datasets, SEA-IFEval (based on IFEval) and SEA-MTBench (based on MT-Bench).

As these two datasets were originally in English, the linguists and native speakers in the team worked together to filter, localise and translate the datasets into the respective target languages to ensure that the examples remained reasonable, meaningful and natural.

SEA-IFEval

SEA-IFEval evaluates a model's ability to adhere to constraints provided in the prompt, for example beginning a response with a specific word/phrase or answering with a certain number of sections. Additionally, accuracy is normalised by the proportion of responses in the correct language (if the model performs the task correctly but responds in the wrong language, it is judged to have failed the task).

SEA-MTBench

SEA-MTBench evaluates a model's ability to engage in multi-turn (2 turns) conversations and respond in ways that align with human needs. We use gpt-4-1106-preview as the judge model and compare against gpt-3.5-turbo-0125 as the baseline model. The metric used is the weighted win rate against the baseline model (i.e. average win rate across each category: Math, Reasoning, STEM, Humanities, Roleplay, Writing, Extraction). A tie is given a score of 0.5.

For more details on Llama-Sahabat-AI-v2-70B-IT benchmark performance, please refer to the Sahabat-AI leaderboard.

Usage

Llama-Sahabat-AI-v2-70B-IT can be run using the 🤗 Transformers library.

To load this model in FP16 or BF16 precision, you need a minimum of approximately 140 GB of VRAM.

Recommended setups include:

4× NVIDIA L40s
2× NVIDIA H100

python

import transformers
import torch
model_id = "Sahabat-AI/Llama-Sahabat-AI-v2-70B-IT"
pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)
messages = [
    {"role": "user", "content": "Sopo wae sing ana ing Punakawan?"},
]
outputs = pipeline(
    messages,
    max_new_tokens=256,
)
print(outputs[0]["generated_text"][-1])

Caveats

It is important for users to be aware that our model exhibits certain limitations that warrant consideration. Like many LLMs, the model can hallucinate and occasionally generates irrelevant content, introducing fictional elements that are not grounded in the provided context. Users should also exercise caution in interpreting and validating the model's responses due to the potential inconsistencies in its reasoning. Please refer to our responsible use guide.

Limitations

Safety

Sahabat-AI-v2-70B-IT has been aligned for general safety topics. However, 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.

Call for Contributions

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.

We also collaborate with key top Indonesia universities such as University of Indonesia, Gadjah Mada University, Bogor Institute of Agriculture, Bandung Institute of Technology, University of North Sumatera (Universitas Sumatera Utara), and Udayana University, including top Indonesian media groups, such as Kompas Gramedia Group, and Republika, Tempo, and Hukumonline 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 Team

PT GoTo Gojek Tokopedia Tbk

Annisa Dininta, Chau Shiau Ching, Saini Ajay Kumar, Shalev Ofir, Tan Daryl, Tep Kilian Rithi, Tiwari Anupam, Widjojo Daniel

AI Singapore

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

University Contributors

Bogor Institute of Agriculture

Hakim Muhammad Luqman, Mashun Hasrul Malik Putra, Muhammad Hariz Krisha, Nurfaza Malikus Syafaadi, Purba Darmawan Setya Putra, Rahardjo Naufal Akbar, Razin Sulthan Farras, Sururi Andika Risky

Udayana University

Agape Febrian Valentino , Jaya I Komang Bisma Bendesa , Juliana Ni Komang Ayu , Putra I Gusti Bagus Sutha Arianata , Putra I Nyoman Adi Mahendra , Widnyana I Gede , Widnyana I Kadek Agus Candra

University of Indonesia

Alsar Muhammad Naufal Zaky , Dharmawan Zanetta Aisha, Malano Naznien Fevrianne, Sinanta Michael Christlambert

University of North Sumatera (Universitas Sumatera Utara)

Naibaho Dewes Agustina, Pasaribu Niken Kirey

Gadjah Mada University

Pasaribu Niken Kirey, Putra Krisna Bayu Dharma, Kosambi Coveeta, Nafis Faris Zaidan, Nailfaaz, Nursanto Achmad Dani, Galih Muhammad Dafa Wisnu

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 the National Research Foundation or the National University of Singapore.

Contact

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

Disclaimer

This is the repository for the Instruct model. While the model has been aligned for general 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.

References

IndoMMLU Reference

bibtex

@inproceedings{koto-etal-2023-large,
    title = "Large Language Models Only Pass Primary School Exams in {I}ndonesia: A Comprehensive Test on {I}ndo{MMLU}",
    author = "Koto, Fajri  and
      Aisyah, Nurul  and
      Li, Haonan  and
      Baldwin, Timothy",
    editor = "Bouamor, Houda  and
      Pino, Juan  and
      Bali, Kalika",
    booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.emnlp-main.760",
    doi = "10.18653/v1/2023.emnlp-main.760",
    pages = "12359--12374",
}