Review

Overlay several runs and decide which one to keep.

Pick two or more jobs and the console overlays their curves, diffs their configs, and shows a scorecard. The question it is built to answer is not “what happened” but “did this get better, and which run do we keep”.

Review: overlaid curves, a scorecard, and the config diff explaining why they differ.

What you see

Region	What it shows
Scorecard	Benchmark scores side by side, with regressions marked by each metric’s declared direction
Primary curves	Validation accuracy and training reward, overlaid
Config diff	Only the keys that actually differ, so a fifty-key config reduces to the three you changed
Ask Agent	Ask in plain language whether the runs improved and which to keep

The config diff is the part people underuse. Two runs that disagree on a curve almost always disagree on a config key, and the diff finds it faster than reading two YAML files.

What you can do

• Add or remove runs — the overlay updates without reloading.
• Ask Agent — natural-language analysis of the selected runs. Locked to administrators unless the deployment sets FORGE_PREMIUM_PUBLIC.
• Share — a read-only link to this exact comparison, with an expiry. See sharing.

Reading a comparison honestly

Identical configs, different curves

Seeds. Post-training runs vary between identical configurations, sometimes considerably. Two runs is not a result; if the difference matters, run the variant again.

Better on the primary metric, worse elsewhere

The normal outcome, and the reason the scorecard shows several benchmarks rather than one. Post- training reliably improves one axis while quietly damaging another — that is what the math, code, knowledge, Chinese and instruction-following split exists to make visible.

Reward climbing, benchmark flat

The model is getting better at your reward, not at the task. Worth looking at the validation samples before celebrating: reward hacking looks exactly like this from the curve alone.