mindchain/rlm-arithmetic-training

GRPO + RLVR Arithmetic Training

Training Qwen3-0.6B-Base on simple arithmetic (2-digit addition/subtraction) using GRPO + RLVR.

Task

Solve arithmetic problems like:

• —47 + 35 = 82
• —92 - 17 = 75

Approach

• —Model: Qwen/Qwen3-0.6B-Base
• —Method: GRPO (Group Relative Policy Optimization) with RLVR (Reinforcement Learning with Verifiable Rewards)
• —Reward: Exact match on answer
• —Steps: 50

Expected Results

Base model (no math training) should perform poorly (<10%), trained model should improve significantly.