typhoon-ai/llama-3-typhoon-v1.5x-70b-instruct

TEXT GENERATIONConcurrent Unit Cost:4Model Size:70BQuant:FP8Context Size:8kTool Calling:SupportedPublished:May 23, 2024License:llama3Architecture:Transformer0.0K Featherless Exclusive Cold

Llama-3-Typhoon-1.5X-70B-instruct is a 70 billion parameter instruct model developed by SCB 10X Typhoon Team, built on Typhoon 1.5 70B and Llama 3 70B Instruct. It is optimized for Thai and English language applications, Retrieval-Augmented Generation (RAG), constrained generation, and reasoning tasks. The model utilizes task-arithmetic model editing for cross-lingual transfer, combining Thai understanding with Llama 3's human alignment. It demonstrates competitive performance with GPT-4-0612 in Thai language benchmarks.

Loading preview...

Model Overview

Llama-3-Typhoon-1.5X-70B-instruct is a 70 billion parameter decoder-only instruct model, developed by the SCB 10X Typhoon Team. It is built upon the Typhoon 1.5 70B and Llama 3 70B Instruct architectures, specifically designed for strong performance in both Thai and English languages.

Key Capabilities & Differentiators

  • Cross-lingual Transfer: Leverages a novel task-arithmetic model editing technique to combine the Thai language understanding of Typhoon with the human alignment of Llama 3 Instruct.
  • Optimized for Applications: Specifically designed for use cases such as Retrieval-Augmented Generation (RAG), constrained generation, and complex reasoning tasks.
  • Competitive Thai Performance: Benchmarks show competitive results against GPT-4-0612 across various Thai-specific evaluations, including ThaiExam, MT-Bench Thai, and IFEval Thai.
  • Agentic Capabilities: Evaluated for agent use-cases, demonstrating strong performance in tasks like GSM8K and HotpotQA in both Thai and English.
  • Model Editing Insight: Research indicates that the most critical features for accurate Thai generation reside in the upper layers of the transformer block, leading to a high ratio of Typhoon components in these backend layers.

Intended Use Cases

This model is particularly well-suited for developers building applications requiring robust Thai and English language processing, especially those involving:

  • RAG systems for information retrieval and generation.
  • Constrained text generation where specific formats or rules must be followed.
  • Complex reasoning tasks requiring logical and structured responses.
  • Agent-based applications that need to interact with various tools and environments.

Finetunes
2 models