RexTRO111/Qwen3-4B-MegaR3ASONER-v1

Model Overview

RexTRO111/Qwen3-4B-MegaR3ASONER-v1 is a 4 billion parameter language model created by merging the MegaR3ASONER LoRA adapter into the Qwen3-4B-Thinking-2507 base model. This full merged Transformers model can be loaded directly without PeftModel and is designed to enhance reasoning capabilities.

Key Capabilities

• Enhanced Reasoning: Specifically fine-tuned for reasoning tasks, leveraging the MegaR3ASONER LoRA.
• Mathematical Problem Solving: Preliminary evaluations on a subset of the gsm8k_cot task showed an 88% flexible extraction exact match and 83% strict match on 100 examples, indicating strong performance in mathematical reasoning.
• Qwen3 Architecture: Built upon the Qwen3-4B-Thinking-2507 base, inheriting its foundational language understanding.

Good For

• Reasoning-intensive applications: Ideal for tasks requiring logical inference, problem-solving, and structured thinking.
• Mathematical tasks: Shows promise in handling mathematical word problems and calculations.
• Developers seeking a specialized reasoning model: Offers a focused approach to improving reasoning over general-purpose models.

Limitations

• May produce incorrect intermediate thoughts before self-correction.
• Can overthink simple prompts, potentially increasing latency and cost for long reasoning traces.
• Preliminary evaluations were limited; full benchmark scores are not available.
• Users should verify answers for high-stakes applications.