JamesX421/SOLID-StepORLM-Qwen3-8B

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 31, 2026Architecture:Transformer Featherless Exclusive Cold

JamesX421/SOLID-StepORLM-Qwen3-8B is an 8-billion parameter Qwen3-based causal language model, fine-tuned using Solver-Informed Self-Distillation (SOLID) and GRPO. It specializes in operations research modeling and generating solver-backed answers, utilizing a COPT-style StepORLM response template. This model is optimized for complex mathematical and optimization problems, demonstrating strong performance on datasets like OptMATH, MAMO-Complex, and InOR. Its primary differentiator is its ability to generate optimization code compatible with COPT environments for execution.

Loading preview...

Overview

JamesX421/SOLID-StepORLM-Qwen3-8B is an 8-billion parameter model built upon the Qwen3 architecture, specifically the Chenyu-Zhou/StepORLM-Qwen3-8B base. It has been fine-tuned using SOLID (Solver-Informed Self-Distillation), incorporating GRPO and solver-informed token-level KL supervision. This model is designed for operations research tasks, focusing on generating solver-backed answers and optimization models.

Key Capabilities

  • Operations Research Modeling: Generates code and solutions for complex optimization problems.
  • Solver-Backed Answer Generation: Produces answers that are verifiable and executable within a solver environment.
  • COPT-Style Response Template: Utilizes a specific response format compatible with COPT (COmmercial Optimization Programming Toolkit) environments.
  • Specialized Training: Benefits from Solver-Informed Self-Distillation and GRPO for enhanced performance in its domain.

Performance Highlights

The model's performance was evaluated by sampling each problem 64 times, using maj@64 for majority-vote accuracy and pass@k for unbiased pass-at-k estimation, with an objective correctness tolerance of 0.001.

  • OptMATH: Achieved maj@64 of 31.33 and pass@1 of 18.25.
  • MAMO-Complex: Demonstrated strong results with maj@64 of 70.44 and pass@1 of 66.43.
  • InOR: Scored maj@64 of 48.00 and pass@1 of 39.81.

Good For

  • Developers and researchers working on operations research problems requiring automated model generation.
  • Applications that need to integrate with COPT or similar optimization solvers.
  • Tasks involving complex mathematical reasoning and optimization code generation.

Finetunes
1 models