ibm-granite/granite-3.3-8b-math-prm-v2

Overview

Granite-3.3-8B-Math-PRM-v2 is an 8-billion parameter model from IBM Research, fine-tuned from Granite-3.3-8B-Instruct. Its primary function is to act as a process reward model (PRM) for mathematical reasoning, assessing the correctness of individual steps in a solution chain. The model is trained on curated, permissively licensed data and released under an Apache 2.0 license.

Key Capabilities

• Mathematical Reasoning Process Supervision: Evaluates the correctness of each step in a mathematical reasoning process, generating 'Y' for correct and 'N' for incorrect. The probability of 'Y' can serve as a numeric reward score.
• Best-of-N Evaluation: Shows strong performance in best-of-N setups for math tasks, outperforming majority voting and other similarly sized PRMs on benchmarks like Math-500.
• Error Identification: Demonstrates robust error detection capabilities on ProcessBench and PRMBench, achieving top-2 performance in its parameter class.
• Code Inference Scaling: Despite its math-centric fine-tuning, the model also exhibits strong inference scaling performance on code benchmarks such as HumanEval and LCBv5.

Intended Use

This model is designed for applications requiring step-by-step validation of mathematical solutions. It takes a question and a response broken into steps, and for each step, it determines if the reasoning chain so far is correct. This functionality is particularly useful for:

• Automated Grading: Providing granular feedback on mathematical problem-solving.
• Reinforcement Learning from Human Feedback (RLHF): Generating reward signals for training other generative models.
• Quality Assurance: Identifying errors in complex reasoning processes generated by other LLMs.