2025 · Public model adaptation

Qwen-0.5B-GRPO: A Fine-Tuned Math Reasoner

A GRPO fine-tune of Qwen2.5-0.5B-Instruct for structured mathematical reasoning using GSM8K.

Research question

Can a small model learn to lay out a maths answer?

Evidence and scope

I adapted Qwen2.5-0.5B-Instruct on GSM8K word problems. The fine-tuning rewards a correct numerical answer and a response that follows the requested structure. These properties are scored separately.

The model card documents a single training epoch, the upstream Qwen model and example inference code. The public weights let other researchers try the adaptation themselves.

Primary source

Contribution
Davut Emre Taşar · Model developer
Source attribution
Davut Emre Taşar
Publication status
Public model adaptation
Publication or platform
Hugging Face Models
Identifier
emre/Qwen-0.5B-GRPO

Sources

Research areas · Language

Related work