Every request below hits a real GPU: base Qwen2.5-1.5B-Instruct and the GRPO-tuned adapter, served side by side from one vLLM process with multi-LoRA. Same prompt, same greedy decoding, same one-shot example the model was trained and evaluated with.
Or run a graded problem:these six have known answers, so they get graded ✓/✗
01Side by side, live
Base Qwen2.5-1.5B-Instruct
reasoning trace
waiting for a problem…
answer
Tuned Qwen2.5-1.5B + GRPO
reasoning trace
waiting for a problem…
answer
enter a problem above
typed problems are ungraded, with no ground truth to check against
Served by vLLM on Modal, scale-to-zero: the container sleeps after a minute idle, so the first request pays a cold start. If the endpoint can't be reached, the six graded problems fall back to their recorded outputs rather than showing nothing. These are live single runs, so a model can land a problem here that it missed in the recorded set on the overview page. The headline 58.8% → 70.0% comes from the full 1,319-problem held-out eval, not from anything on this page.