We built an RL environment for credit card reward optimization and trained Qwen 32B with GRPO against it. The trained model scores ~0.51 on held-out tasks vs. Opus 4 at ~0.41 and GPT-4o at 0.36. Environment is open source (Apache 2.0). Blog post explains the reward design, what broke during training, how we fixed it, and what we'd do differently.
Show HN: We trained a 32B model to beat Opus 4 at credit card optimization | Heykuki News