begawansemar7/Llama-Sahabat-AI-v2-70B-IT
Llama-Sahabat-AI-v2-70B-IT is a 70 billion parameter decoder-only large language model, co-initiated by PT GoTo Gojek Tokopedia Tbk and Indosat Ooredoo Hutchison, and developed by PT GoTo Gojek Tokopedia Tbk and AI Singapore. This model is instruct-tuned for Indonesian and supports multiple local languages including Javanese, Sundanese, Batak Toba, and Balinese, in addition to English. With a context length of 128k tokens, it is specifically designed to excel in tasks requiring understanding and generation in the Indonesian linguistic and cultural context.
Loading preview...
Llama-Sahabat-AI-v2-70B-IT: Indonesian-Focused Multilingual LLM
Llama-Sahabat-AI-v2-70B-IT is a 70 billion parameter instruction-tuned large language model developed by PT GoTo Gojek Tokopedia Tbk and AI Singapore, with co-initiation from Indosat Ooredoo Hutchison. It is part of the Sahabat-AI collection, specifically designed and optimized for the Indonesian language and its diverse local dialects. The model utilizes the default Llama 3.1 70B Instruct tokenizer and features an extended context length of 128k tokens.
Key Capabilities & Evaluation
This model has been rigorously evaluated on its general language and instruction-following capabilities, with a strong emphasis on the Indonesian context:
- Multilingual Support: Supports English, Indonesian, Javanese, Sundanese, Batak Toba, and Balinese.
- Indonesian Contextual Understanding: Evaluated using the IndoMMLU benchmark across various subjects like Humanities, Indonesian language, local languages and cultures, social science, and STEM for different educational levels.
- General Language Tasks: Assessed on the SEA-HELM benchmark for tasks such as Question Answering, Sentiment Analysis, Toxicity Detection, Translation, Abstractive Summarization, Causal Reasoning, and Natural Language Inference.
- Instruction Following: Performance on instruction adherence is measured with SEA-IFEval (based on IFEval), and multi-turn conversational abilities are evaluated using SEA-MTBench (based on MT-Bench), with datasets localized and translated by linguists.
Usage Considerations
Running this 70B parameter model requires substantial computational resources, specifically a minimum of approximately 140 GB of VRAM for FP16 or BF16 precision. Recommended setups include 4× NVIDIA L40s or 2× NVIDIA H100 GPUs. Users should be aware of common LLM limitations such as potential hallucination and occasional irrelevant content generation. The Sahabat-AI ecosystem is an open-source initiative inviting contributions from researchers and developers to enhance its capabilities and expand its reach within the Indonesian linguistic landscape.