nics-efc/VPR-Qwen3-4B-Sokoban

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 18, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The nics-efc/VPR-Qwen3-4B-Sokoban model is a 4 billion parameter Qwen3 checkpoint, fine-tuned using Verifiable Process Rewards (VPR) on Markovian Sokoban interactions. This model is specifically optimized for agentic reasoning within the Sokoban environment, achieving a 28.40% success rate. It is designed for task-grounded problem-solving rather than general-purpose assistance, leveraging a search-based Sokoban oracle for reward generation.

Loading preview...

Model Overview

The nics-efc/VPR-Qwen3-4B-Sokoban is a 4 billion parameter model based on the Qwen3 architecture, developed by nics-efc. It has been specifically trained using Verifiable Process Rewards (VPR) on Markovian Sokoban interactions, focusing on agentic reasoning within this particular game environment. The training process involves sampling action responses, scoring them with a task-grounded search-based Sokoban oracle, and optimizing eligible candidates using locally normalized advantages.

Key Capabilities

  • Specialized Agentic Reasoning: Excels at solving Sokoban puzzles by processing Markovian game prompts and generating actions.
  • VPR Training: Utilizes a unique reward mechanism that samples and scores actions with an oracle, committing high-reward candidates.
  • Performance in Sokoban: Achieves a reported success rate (SR) of 28.40% (± 2.79) on the Sokoban task under the VPR evaluation protocol.

Use Cases and Limitations

This model is not intended as a general-purpose assistant. Its primary application is in research and development related to agentic reasoning within highly structured, task-grounded environments like Sokoban. Users should leverage the environment prompts, parsers, and action conventions from the VPR codebase for reproduction and optimal performance. The model's training relies on specific task-grounded oracle signals and Markovian prompts, meaning its performance outside these documented environments and action formats has not been established. Safety and correctness should be evaluated before any open-ended deployment.

Finetunes
1 models