Technical Architecture for Reasoning Capabilities of Seed-OSS
The Seed-OSS-36B model utilizes the following innovative designs to achieve powerful reasoning capabilities:
- mathematical reasoning: 40% accuracy improvement in AIME benchmarks
- code generation: LiveCodeBench evaluation shows that its problem solving efficiency is ahead of its class 15%
- Dynamic budgetary control: The thinking_budget parameter allows the developer to adjust the depth of reasoning as required
The model training uses only 12 trillion tokens to achieve the current performance, proving the efficiency of its algorithm. Practical applications support the automatic tool invocation function to automate the processing of computation, query and other agent tasks.
This answer comes from the articleSeed-OSS: Open Source Large Language Model for Long Context Reasoning and Versatile ApplicationsThe
May not be reproduced without permission:AI productivity tools " Seed-OSS's reasoning power is optimized to excel on complex tasks
