NuminaMath 7B TIR Released: Transforming Mathematical Problem-Solving with Advanced Tool-Integrated Reasoning and Python REPL for Competition-Level Accuracy
What is the total parameter count of NuminaMath 7B TIR?
The NuminaMath 7B TIR model has a total parameter count of 6.91 billion. This advanced language model is designed specifically for solving mathematical problems and is adept at handling complex mathematical queries through a sophisticated tool-integrated reasoning (TIR) mechanism3.
How does NuminaMath 7B TIR execute Python code?
NuminaMath 7B TIR executes Python code through a structured process. It first generates a detailed reasoning pathway to approach the problem, then translates this reasoning into executable Python code2. The Python code is executed in a REPL (Read-Eval-Print Loop) environment. If the initial attempt fails, the model attempts to self-heal by iterating through the steps using the incorrect output until a correct solution is found.
What inspired the second fine-tuning stage of NuminaMath 7B TIR?
The second fine-tuning stage of NuminaMath 7B TIR was inspired by Microsoft's ToRA (Tool-integrated Reasoning Agent) framework5. This stage focused on a synthetic dataset emphasizing tool-integrated reasoning, decomposing each math problem into a sequence of rationales, Python programs, and their outputs. This approach aimed to enhance the model's ability to solve mathematical problems by combining natural language reasoning with computational tools.