pngwn/qwen2.5-0.5b-gsm8k-sft

TEXT GENERATIONConcurrency Cost:1Model Size:0.5BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Jun 10, 2026Architecture:Transformer Cold

The pngwn/qwen2.5-0.5b-gsm8k-sft model is a 0.5 billion parameter Qwen2.5-based causal language model, specifically supervised fine-tuned for grade-school mathematical reasoning. Developed by pngwn, it demonstrates a significant improvement in GSM8K test exact-match accuracy compared to its base model. This model is optimized for solving arithmetic word problems and is suitable for applications requiring focused mathematical problem-solving capabilities.

Loading preview...

Model Overview

The pngwn/qwen2.5-0.5b-gsm8k-sft is a 0.5 billion parameter model built upon the Qwen2.5 architecture. Its primary distinction lies in its supervised fine-tuning (SFT) specifically for grade-school mathematical reasoning using the openai/gsm8k dataset.

Key Capabilities & Performance

This model excels at solving arithmetic word problems, as evidenced by its performance on the GSM8K benchmark. It achieves an exact-match accuracy of 0.3472 (458/1319) on the GSM8K test set, a substantial improvement over the base Qwen/Qwen2.5-0.5B model's accuracy of 0.0008. This specialization makes it highly effective for tasks requiring precise numerical problem-solving.

Training Details

The model was trained for 3 epochs on 7473 samples from the openai/gsm8k dataset, with a maximum sequence length of 1024 tokens. Decoding was performed greedily, and answers were extracted using a specific regex pattern.

Use Cases

This model is particularly well-suited for:

  • Educational applications: Assisting students with math homework or generating practice problems.
  • Automated grading: Evaluating solutions to mathematical problems.
  • Specialized reasoning tasks: Any application where accurate, step-by-step mathematical problem-solving is critical, especially at a foundational level.