typhoon-ai/llama-3-typhoon-v1.5x-8b-instruct
Llama-3-Typhoon-1.5X-8B-instruct is an 8 billion parameter instruction-tuned model developed by SCB 10X Typhoon Team, built on Typhoon 1.5 8B and Llama 3 8B Instruct. This model is optimized for Thai and English languages, excelling in application use cases, Retrieval-Augmented Generation (RAG), constrained generation, and reasoning tasks. It leverages cross-lingual transfer and task-arithmetic model editing to combine Thai understanding with human alignment, demonstrating competitive performance against GPT-3.5-turbo in specific benchmarks.
Loading preview...
Llama-3-Typhoon-1.5X-8B-instruct: Thai-Optimized LLM
This 8 billion parameter instruct model, developed by the SCB 10X Typhoon Team, is a hybrid built upon Typhoon 1.5 8B and Llama 3 8B Instruct. It leverages cross-lingual transfer and task-arithmetic model editing to merge Typhoon's Thai language understanding with Llama 3's human alignment capabilities. The model is primarily designed for Thai and English languages, supporting an 8192-token context length.
Key Capabilities & Performance
- Bilingual Proficiency: Strong performance in both Thai and English, with a focus on Thai language tasks.
- Optimized for Applications: Specifically designed for practical use cases including Retrieval-Augmented Generation (RAG), constrained generation, and complex reasoning.
- Competitive Benchmarks: Demonstrates performance comparable to GPT-3.5-turbo in several evaluations:
- ThaiExam: Achieves an average of 0.5028, matching Typhoon-1.5 8B and outperforming gpt-3.5-turbo-0125.
- MMLU: Scores 0.6369, showing improvement over Typhoon-1.5 8B.
- MT-Bench Thai: Scores 6.902, surpassing both Typhoon-1.5 8B and gpt-3.5-turbo-0125 in Thai human alignment.
- IFEval Thai: Scores 0.548, outperforming gpt-3.5-turbo-0125 in Thai instruction following.
- Model Editing Insight: Utilizes model editing techniques, incorporating a high ratio of Typhoon components in the upper transformer layers to enhance Thai answer generation accuracy.
Good For
- Thai Language Applications: Ideal for developers building applications requiring robust Thai language understanding and generation.
- RAG Systems: Well-suited for integrating with Retrieval-Augmented Generation pipelines.
- Constrained Generation: Excels at tasks requiring adherence to specific output formats or rules.
- Reasoning Tasks: Capable of generating clear and logically structured responses across multiple steps.