AB
AB
AiBoss

All Categories

project

MetaStone-L1-7B - A lightweight inference model launched by MetaStone Intelligent Computing

MetaStone-L1-7B is a lightweight inference model in the MetaStone series, designed to improve performance for complex downstream tasks. It achieves state-of-the-art performance among parallel models in core inference benchmarks such as math and code (SO...

What is MetaStone-L1-7B?

MetaStone-L1-7B is a lightweight inference model in the MetaStone series, designed to improve performance for complex downstream tasks. It achieves state-of-the-art (SOTA) performance for parallel models in core inference benchmarks such as math and code, comparable to API models like Claude-3.5-Sonnet-1022 and GPT4o-0513. The model is trained on DeepSeek-R1-Distill-Qwen-7B using GRPO.

Main functions of MetaStone-L1-7B

  • Strong reasoning abilityIt achieves state-of-the-art (SOTA) performance in core reasoning benchmarks such as math and code, comparable to API models like Claude-3.5-Sonnet-1022 and GPT4o-0513.
  • Optimized training foundationIt is based on DeepSeek-R1-Distill-Qwen-7B and trained by GRPO, thus possessing a good performance foundation.
  • Flexible use of settingsFor optimal performance, it is recommended to use a temperature of 0.6°C, a top sampling probability of 0.95, and a maximum generation length of 32k.
  • Optimization tips for specific problemsFor math problems, add "" to the prompt.Please reason step by step, and put your final answer within \\boxed{}.For code issues, adding specific formatting requirements to the prompts can further improve the model's inference performance.

Technical Principles of MetaStone-L1-7B

  • Technical architecture based on DeepSeek-R1MetaStone-L1-7B is trained on DeepSeek-R1-Distill-Qwen-7B using GRPO. DeepSeek-R1 itself employs advanced Visual Fine-Tuning (V-FT) technology, significantly improving the model's visual understanding and multimodal alignment capabilities through innovative methods such as cross-modal contrastive learning, gradient decoupling training, and attention gating mechanisms.
  • Support for cross-architecture computing clustersMetaStone has proposed a RISC-V & x86 cross-architecture computing cluster solution. Through a "one cloud, multiple chips" design, it is compatible with various CPU chip architectures, including RISC-V and x86. This enables the MetaStone-L1-7B to run efficiently on computing resources of different architectures, fully leveraging the performance advantages of CPUs with different architectures.
  • Application of cloud-native technologiesMetaStone has innovatively proposed "MetaStone CloudOS," a cloud-native technology platform that provides applications with full cloud infrastructure capabilities, including containers, virtualization, storage, networking, and security. MetaStone CloudOS breaks down the barriers in resource scheduling between RISC-V and x86/ARM architecture clusters, enabling true cross-architecture computing power flow.
  • Low-parameter, high-efficiency designMetaStone-L1-7B requires only 7B parameters to achieve the same text-image understanding capabilities as traditional multimodal models (such as Flamingo-80B). This low parameter count design reduces computational resource requirements and further improves training efficiency through techniques such as mixed-precision training and gradient checkpointing.

MetaStone-L1-7B project address

Application scenarios of MetaStone-L1-7B

  • Mathematical Problem SolvingIt can logically solve complex mathematical problems step by step, such as squaring or solving equations. Users simply input the problem, and the model will logically reason step by step and provide the final answer in a specific format (e.g., ...). \boxed{}Presented in a way that facilitates student understanding and learning.
  • Programming aidsFor programming problems, the model can generate code that meets the requirements and output it in a specified format. For example, a user can ask the model to read the input, solve the problem, and write the answer to the output; the model will then generate the corresponding code snippets.
  • Intelligent Customer ServiceIt can quickly and accurately answer users' questions, provide solutions and suggestions, and improve the efficiency and quality of customer service.
  • Content creationIt helps users generate text content, such as articles, stories, and poems, and inspires creative ideas.
  • Code generation and optimizationIt generates corresponding code snippets based on user needs, helping developers quickly implement functions and improve development efficiency.
Back