LLMOPT-Qwen2.5-14B: Optimization Problem Solver

LLMOPT-Qwen2.5-14B is a 14.8 billion parameter model, developed by Ant Group, East China Normal University, and Nanjing University, specifically fine-tuned from Qwen2.5-14B-Instruct to address general optimization problems. This model excels at interpreting natural language descriptions of optimization tasks and generating executable solutions.

Key Capabilities

• High Solving Accuracy: Achieves an average Solving Accuracy (SA) of 76.40% with self-correction across diverse optimization datasets, including NL4Opt, Mamo, NLP4LP, and IndustryOR.
• Robust Execution Rate: Demonstrates an average Execution Rate (ER) of 94.48% with self-correction, indicating its ability to produce runnable code for solutions.
• Self-Correction Mechanism: Utilizes a self-correction process to refine solutions, significantly improving both execution rate and solving accuracy compared to initial attempts.
• Specialized Dataset Training: Evaluated and trained on a comprehensive suite of optimization-specific datasets, including manually curated and re-labeled data from sources like NL4Opt, Mamo, NLP4LP, ComplexOR, IndustryOR, ICML Competition, OptiBench, and OptMath.

Good For

• Automated Optimization Modeling: Ideal for applications requiring the automatic definition and solution of complex optimization problems from textual input.
• Operational Research: Suitable for tasks in operational research, logistics, resource allocation, and other domains where optimization is critical.
• Research and Development: Provides a strong baseline for further research into LLM-based optimization and problem-solving.