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

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:Transformer Open Weights Featherless Exclusive Cold

The nics-efc/VPR-Qwen3-4B-Minesweeper model is a 4 billion parameter Qwen3 checkpoint developed by nics-efc. It is specifically trained with Verifiable Process Rewards (VPR) on Markovian Minesweeper interactions. This model excels at agentic reasoning within the Minesweeper environment, demonstrating a 32.60% success rate and 85.76% completion rate. It is optimized for task-grounded oracle signals and specific Markovian prompts, making it suitable for research in verifiable process rewards.

Loading preview...

Model Overview

The nics-efc/VPR-Qwen3-4B-Minesweeper is a 4 billion parameter Qwen3 model developed by nics-efc. It has been specifically fine-tuned using Verifiable Process Rewards (VPR) on interactions within a Markovian Minesweeper environment. This training methodology involves sampling action responses, scoring them with a task-grounded posterior-based Minesweeper oracle, and optimizing eligible candidates using locally normalized advantages.

Key Capabilities

  • Specialized Agentic Reasoning: Designed for agentic reasoning within the specific context of Markovian Minesweeper.
  • VPR Training: Utilizes a unique Verifiable Process Rewards mechanism for training, focusing on verifiable and task-grounded oracle signals.
  • Performance in Minesweeper: Achieves a reported 32.60% success rate (SR) and 85.76% completion rate (CR) on the Minesweeper evaluation protocol defined in the VPR paper.

Limitations and Intended Use

This model is a task-specific checkpoint and not intended as a general-purpose assistant. Its training relies heavily on task-grounded oracle signals and specific Markovian prompts. Performance outside these documented environments and action formats has not been established. Users should evaluate its safety and correctness thoroughly before any open-ended deployment. For reproduction and further details, refer to the VPR codebase and the associated paper.