The Evals page lets you inspect evaluation results at each step where an eval was run. It uses the same sample viewer as the Rollouts page, adapted for eval data.
You can switch between different eval environments using the dropdown in the left sidebar. Each eval shows its samples organized by step — navigate between steps to see how the model’s eval performance evolves during training.For evals configured with multiple completions per prompt (pass@k), clicking a sample expands to show all completions so you can compare them.
Each eval sample shows the prompt, the model’s completion, per-metric scores, and golden answers when available. The same render options (think blocks, markdown, LaTeX, code) are available for inspecting outputs.
The right sidebar shows eval-specific metric charts over training steps, so you can track how eval scores trend alongside the samples you’re inspecting.
