AB
AB
AiBoss

All Categories

project

FoxBrain - A large-scale reasoning language model launched by Hon Hai Research Institute

FoxBrain is a large-scale Traditional Chinese language model (LLM) developed by Foxconn Research Institute, boasting powerful reasoning capabilities. Based on the Meta Llama 3.1 architecture, it has 70 parameters and focuses on mathematical and logical reasoning. FoxB...

What is FoxBrain?

FoxBrain is a large-scale Traditional Chinese language model (LLM) developed by Foxconn Research Institute, boasting powerful reasoning capabilities. Based on the Meta Llama 3.1 architecture, it has 70 parameters and focuses on mathematical and logical reasoning. FoxBrain utilizes an efficient training strategy, completing training in four weeks using 120 NVIDIA H100 GPUs, demonstrating high efficiency and low cost. FoxBrain generates high-quality Chinese pre-training data using proprietary technology and enhances its reasoning capabilities with Adaptive Reasoning Reflection. FoxBrain is driving the application of AI in manufacturing, supply chain management, and intelligent decision-making, setting a new benchmark for AI technology development in Taiwan.

FoxBrain's main functions

  • Data Analysis and Decision SupportIt processes complex data and provides decision support for users.
  • Document CollaborationIt helps users generate, edit, and optimize text content, improving work efficiency.
  • Mathematics and Reasoning Problem SolvingThey excel in mathematical and logical reasoning, solving complex mathematical and logical problems.
  • Code generationIt helps users generate programming code and improves development efficiency.
  • Language style optimizationThe language style of Traditional Chinese in Taiwan has been optimized to better suit the needs of local users.

FoxBrain's technical principles

  • InfrastructureBased on the Meta Llama 3.1 architecture, it has 70B parameters and powerful language generation and understanding capabilities.
  • High-efficiency training strategiesTraining was performed using 120 NVIDIA H100 GPUs, based on a multi-node parallel training architecture to ensure high efficiency and stability. High-quality Chinese pre-training data with 98 bytes of tokens, covering 24 topics, was generated using proprietary technology. Adaptive Reasoning Reflection technology was used to train the model's autonomous reasoning ability, improving its performance in mathematical and logical reasoning domains.
  • Context processing capabilitiesSupports context processing lengths of up to 128K tokens, enabling the handling of longer text inputs and outputs.
  • Optimization and EvaluationBased on techniques such as Continual Pre-Training, Supervised Fine-tuning, and RLAIF, the model performance is continuously optimized, and it performs excellently on the TMMLU+ test dataset.

FoxBrain's project address

Application scenarios of FoxBrain

  • Smart manufacturing and supply chain managementOptimize production processes, provide data analysis and decision support, and improve supply chain efficiency and intelligence.
  • Smart Education and Learning AssistanceIt helps students solve math and logic problems, provides learning guidance, generates learning materials, and improves learning efficiency.
  • Smart office and document collaborationIt automatically generates and optimizes text content, assisting in writing reports, emails, etc., thereby improving office efficiency and document quality.
  • Software Development and Code GenerationIt helps programmers generate code snippets, provides programming suggestions, and improves software development efficiency and quality.
  • Smart Cities and Data AnalyticsIt supports the analysis and processing of urban data, provides intelligent decision support, and helps to upgrade urban management and public services to a smarter level.
Back