AI BriefingUpstage 13B LLM
Key point
Upstage-Qanda 13B LLM achieved SoTA on GSM8K and MATH.
Details
Upstage-Qanda 13B LLM achieved SoTA in the math problem-solving domain. It surpassed MS ToRA 13B and ChatGPT, and on MATH it recorded 48.8% accuracy, higher than GPT-4.
Upstage collaborated with Qanda and KT to build a model specialized for math reasoning. Based on Qanda's high-quality math data and GPUs provided by KT, it combined natural language reasoning with program-based calculation.
The key methods are as follows.
- Chain-of-Thought: strengthens natural language reasoning
- Program-of-Thought: supplements precise calculation
- Achieved SoTA on both GSM8K and MATH
Upstage emphasized that this achievement demonstrates the competitiveness of domain-specific LLMs. It stated that it will expand into more diverse global AI use cases centered on its flagship model Solar going forward, and the research methods and analysis will be released in a follow-up paper.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.