JamesX421/SOLID-opt
JamesX421/SOLID-opt is a 4 billion parameter Qwen3-4B-Instruct-2507 checkpoint specifically fine-tuned for operations research modeling. This model excels at transforming natural language optimization problems into structured reasoning, linear programming formulations, and solver-oriented Python code. It is designed for research in mathematical optimization, providing specialized capabilities for generating optimization problem solutions. The model leverages a unique training approach combining GRPO, solver-objective majority voting, and teacher KL localized with solver-information signals.
Loading preview...
SOLID-opt: Specialized for Operations Research
SOLID-opt is a 4 billion parameter model based on the Qwen3-4B-Instruct-2507 architecture, developed by JamesX421. Its core specialization lies in operations research (OR) modeling, aiming to bridge the gap between natural language problem descriptions and executable optimization solutions.
Key Capabilities
- Natural Language to Optimization: Translates optimization problems described in natural language into structured formats.
- Formulation Generation: Capable of generating linear programming formulations from problem statements.
- Solver-Oriented Code: Produces Python code designed to interact with optimization solvers.
- Structured Reasoning: Facilitates structured reasoning for complex optimization challenges.
Training Methodology
The model's unique capabilities stem from its specialized training regimen, which includes:
- GRPO (Generalized Policy Optimization): A reinforcement learning technique.
- Solver-Objective Majority Voting: Incorporates feedback based on solver objectives.
- Teacher KL Localized with Solver-Information Signals: Utilizes knowledge distillation with solver-specific information.
This training configuration employs Gurobi-style nine-step outputs and a teacher-KL coefficient of 0.001.
Intended Use and Limitations
SOLID-opt is released primarily for research in mathematical optimization and operations-research reasoning. Users should be aware that generated formulations, coefficients, constraints, solver code, and claimed solutions may be incorrect, infeasible, or unsafe. It is crucial to validate all outputs with an appropriate solver and independent checks before deployment in any consequential setting. No standalone evaluation results are provided with this initial release.