KRX-Data/WON-Reasoning

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Mar 20, 2025Architecture:Transformer0.0K Warm

KRX-Data's WON-Reasoning is a 7.6 billion parameter, Qwen2.5-Math-7B-Instruct based large language model with a 131072 token context length, specifically fine-tuned for Korean financial tasks. It employs a two-step structured reasoning approach, providing self-correcting reasoning and conclusive summaries, and excels in financial reasoning, accounting, and open-ended financial question-answering.

Loading preview...

₩ON: Korean Financial LLM

₩ON is an advanced 7.6 billion parameter Large Language Model (LLM) developed by KRX-Data, specifically designed for financial tasks within the Korean domain. Built on the Qwen2.5-Math-7B-Instruct base model, it features a substantial 131072 token context length and supports both Korean and English.

Key Capabilities

  • Specialized Financial Reasoning: Tailored for complex financial analysis, accounting principles, and econometric reasoning.
  • Structured Output: Employs a unique two-step reasoning process, outputting a detailed thought process within <think> tags followed by a concise conclusion in <solution> tags, enhancing transparency and clarity.
  • High Performance: Achieved the highest average performance in the KRX Financial LLM Competition benchmark, particularly excelling in Finance & Accounting and Open-Ended FinQA tasks.
  • Robust Training: Utilizes a sophisticated two-phase training strategy involving Supervised Fine-Tuning (SFT) on a 400,000-sample dataset and Direct Preference Optimization (DPO) to refine responses and reduce overthinking.

Good For

  • Developers and researchers building AI applications for Korean financial markets.
  • Tasks requiring rigorous financial reasoning and transparent, structured answers.
  • Analyzing financial documents, market data, and accounting principles in a Korean context.