Model

Qwen 0.5B · GRPO maths

Purpose

Can a small model learn both to solve a maths word problem and to lay out its answer? This half-billion-parameter Qwen model trained for one pass over GSM8K. It writes out a solution and places the final answer in a separate section. During GRPO training, custom rewards checked the numeric answer and the response format separately. A neatly formatted solution can still get the arithmetic wrong.

Base model or method
Qwen2.5-0.5B-Instruct

Sources

Qwen 0.5B · GRPO maths