HuggingFaceTB/qwen3-1.7b-gsm8k-sft
HuggingFaceTB/qwen3-1.7b-gsm8k-sft is a 1.7 billion parameter Qwen3-based causal language model fine-tuned specifically for mathematical reasoning. It achieves 77.2% accuracy on the GSM8K benchmark, a significant improvement over its base model, and also performs well on competition-level math problems. This model is optimized for solving grade school math word problems using chain-of-thought reasoning, making it suitable for applications requiring robust numerical problem-solving capabilities.
Loading preview...
Qwen3-1.7B Fine-tuned for GSM8K Math Reasoning
This model is a specialized fine-tuned version of the Qwen3-1.7B base model, developed by HuggingFaceTB, with a primary focus on enhancing mathematical reasoning capabilities. It has been specifically optimized for performance on the GSM8K (Grade School Math) benchmark, demonstrating a substantial improvement in accuracy compared to its pre-trained counterpart.
Key Capabilities
- Exceptional GSM8K Performance: Achieves 77.2% accuracy on the GSM8K benchmark, a significant increase of over 57 percentage points from the base model's ~20%.
- Chain-of-Thought Reasoning: Utilizes a
<think>...</think>chain-of-thought format during inference, which is crucial for its improved math problem-solving. - Competition Math Proficiency: Shows solid performance on the MATH-500 benchmark (55.2% accuracy), particularly strong in Algebra (71.8%) and Prealgebra (68.3%).
- Efficient Deployment: Available in GGUF quantized versions (e.g., 8-bit q8_0.gguf at 1.8 GB) for efficient deployment with
llama.cppand Ollama.
Good for
- Mathematical Problem Solving: Ideal for applications requiring accurate solutions to grade school-level math word problems.
- Educational Tools: Can be integrated into tutoring systems or educational platforms that need to generate step-by-step math solutions.
- Reasoning Tasks: Suitable for tasks where explicit, structured reasoning (like chain-of-thought) is beneficial for achieving correct answers.
- Resource-Constrained Environments: Its 1.7 billion parameter size and availability of quantized versions make it suitable for deployment on devices with limited computational resources.