Lzvick/qwen-1.7b-math-reasoner-grpo
The Lzvick/qwen-1.7b-math-reasoner-grpo is a 2 billion parameter Qwen3-based language model fine-tuned using Group Relative Policy Optimization (GRPO) on the GSM8K dataset. It is specifically optimized for mathematical reasoning and solving grade-school arithmetic word problems. This model demonstrates significant improvements in math accuracy while retaining general knowledge capabilities, making it suitable for specialized numerical tasks.
Loading preview...
Model Overview
Lzvick/qwen-1.7b-math-reasoner-grpo is a 2 billion parameter model based on unsloth/Qwen3-1.7B-Base, fine-tuned using Group Relative Policy Optimization (GRPO). This model was trained with verifiable rewards (RLVR) on the GSM8K math word problem dataset for one epoch, focusing on improving mathematical reasoning capabilities.
Key Capabilities & Performance
- Enhanced Mathematical Reasoning: Achieved an 87.3% accuracy on held-out GSM8K examples, a +20.4 percentage point improvement over its base model. It also showed significant gains on out-of-distribution datasets like SVAMP (+10.0pp) and MATH/algebra (+11.3pp).
- Robustness to Catastrophic Forgetting: Independent cross-checks using
lm-evaluation-harnessconfirmed substantial GSM8K improvements without degradation in general capabilities (MMLU, HellaSwag scores remained stable). - Structured Output: Trained to produce responses in a fixed
<reasoning> ... </reasoning> <answer> ... </answer>format, with 99.4% compliance. - Reward System: Utilizes a reward system prioritizing correctness, format compliance, and brevity.
Limitations & Considerations
- Narrow Specialization: Primarily trained on grade-school arithmetic; general instruction-following and chat quality were not the focus.
- Hint Exploitation Vulnerability: Exhibits a 37% vulnerability to following plausible-but-wrong hints in prompts, even when its own reasoning contradicts them. Users should avoid presenting untrusted "hint" text.
- Approximate MATH Evaluation: MATH benchmark scores are directional due to grading complexities.
- Output Stopping: The model may occasionally generate content past the
</answer>tag; it is recommended to set a stop sequence during generation (e.g.,stop=["</answer>"]). - Weak Spot in Symbolic Function Composition: Anecdotal evidence suggests difficulty with abstract symbolic function notation (e.g.,
f(x)=...).
Recommended Use Cases
This model is ideal for applications requiring strong mathematical reasoning, particularly for solving arithmetic word problems. Its specialized training makes it a powerful tool for tasks where numerical accuracy and structured output are critical, especially within the domain of grade-school mathematics.