nics-efc/VPR-Qwen3-4B-Sudoku
The nics-efc/VPR-Qwen3-4B-Sudoku is a 4 billion parameter Qwen3 checkpoint developed by nics-efc, specifically trained with Verifiable Process Rewards (VPR) for Markovian Sudoku interactions. This model excels at solving Sudoku puzzles, achieving an 80.60% success rate and 84.22% completion rate by optimizing action responses with a task-grounded constraint-based oracle. It is designed for specialized agentic reasoning tasks rather than general-purpose assistance.
Loading preview...
VPR-Qwen3-4B-Sudoku: Specialized Sudoku Solver
This model is a 4 billion parameter Qwen3 checkpoint, developed by nics-efc, uniquely trained using Verifiable Process Rewards (VPR). Its core innovation lies in its training methodology, which involves sampling multiple action responses for Markovian Sudoku interactions, scoring them with a constraint-based Sudoku oracle, and then optimizing eligible candidates based on locally normalized advantages.
Key Capabilities
- Specialized Sudoku Solving: Achieves a reported 80.60% success rate (SR) and 84.22% completion rate (CR) on Sudoku tasks, as evaluated under the VPR paper's protocol.
- Constraint-Based Optimization: Utilizes a task-grounded Sudoku oracle to verify and select optimal actions, ensuring high accuracy in puzzle resolution.
- Agentic Reasoning: Designed for agentic reasoning within specific, structured environments like Sudoku, where actions can be verified against clear rules.
Limitations and Intended Use
This model is not intended as a general-purpose assistant. Its training is highly dependent on task-grounded oracle signals and specific Markovian prompts. Performance outside of the documented environments and action formats has not been established. Users should refer to the VPR codebase for environment prompts, parsers, and action conventions necessary for reproduction and proper usage. Evaluate safety and correctness carefully before any open-ended deployment.