JamesX421/ttrl-opt
JamesX421/ttrl-opt is a 4 billion parameter Qwen3-4B-Instruct-2507 checkpoint, specialized for operations-research modeling with a 32768 token context length. It serves as a baseline for generating structured reasoning, linear-programming formulations, and solver-oriented Python code from natural-language optimization problems. This model was trained using GRPO and TTRL-style majority voting with Gurobi-oriented optimization outputs and solver-based reward signals. Its primary strength lies in assisting with mathematical optimization and operations-research reasoning tasks.
Loading preview...
Overview
JamesX421/ttrl-opt is a specialized 4 billion parameter model, based on the Qwen3-4B-Instruct-2507 architecture, designed for operations-research modeling. It focuses on generating structured reasoning, linear-programming formulations, and Python code tailored for solvers from natural-language optimization problems. This particular release is a checkpoint from training step 125, developed using GRPO and TTRL-style majority voting, incorporating Gurobi-oriented optimization outputs and solver-based reward signals.
Key Capabilities
- Operations Research Modeling: Generates formulations and code for optimization problems.
- Structured Reasoning: Produces structured reasoning outputs for complex problems.
- Solver-Oriented Code: Creates Python code specifically for optimization solvers.
- Linear Programming: Formulates linear programming problems from natural language.
Intended Use and Limitations
This model is released for research purposes in mathematical optimization and operations-research reasoning. Users should be aware that generated formulations, coefficients, constraints, solver code, and proposed solutions may contain errors, be infeasible, or unsafe without thorough review. It is crucial to validate all outputs with an appropriate solver and independent checks before deployment in any consequential settings. As of this initial release, no standalone evaluation results are provided.