cesun/ReFIne-qwen3-1.7b

Hugging Face
TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 10, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

The cesun/ReFIne-qwen3-1.7b model is a 1.7 billion parameter Qwen3-based large reasoning model developed by cesun, part of the ReFIne framework. This model is specifically fine-tuned to enhance interpretability, faithfulness, and reliability in its reasoning processes, particularly for mathematical benchmarks. It achieves this by generating structured, tag-based traces, disclosing decisive information, and providing self-assessments of its derivations and confidence. Its primary use case is in applications requiring transparent and trustworthy reasoning, especially in mathematical problem-solving.

Loading preview...

ReFIne-qwen3-1.7b: Trustworthy Large Reasoning Models

The ReFIne-qwen3-1.7b model, developed by cesun, is a 1.7 billion parameter model based on the Qwen3 architecture. It is a key component of the ReFIne framework, which aims to significantly improve the interpretability, faithfulness, and reliability of Large Reasoning Models (LRMs). This framework integrates supervised fine-tuning with GRPO to guide models in producing more transparent and verifiable reasoning.

Key Capabilities

  • Enhanced Interpretability: Generates structured, tag-based reasoning traces with high-level planning, making the model's thought process easier for humans to follow.
  • Improved Faithfulness: Explicitly identifies and discloses the decisive information used at each step of a solution, ensuring consistency and traceability.
  • Increased Reliability: Provides self-assessments of the soundness of its derivations and the confidence level of its final answers, fostering greater trust in its outputs.
  • Mathematical Reasoning: Demonstrated significant improvements in trustworthiness dimensions when evaluated across mathematical benchmarks.

Good for

  • Applications requiring transparent and verifiable reasoning.
  • Mathematical problem-solving where step-by-step interpretability is crucial.
  • Developing AI systems that need to explain their decisions and assess their own confidence.
  • Research into trustworthy AI and explainable AI (XAI) for reasoning tasks.

For more detailed information, including installation and training procedures, refer to the official GitHub repository.

Finetunes
1 models