extractingbench

Confidence you can trust

When it says 90% sure, is it right 90% of the time?

Only for models that return a confidence. Others show not supported.

For values the model has to estimate, we check whether its 80% range contains the truth about 80% of the time.

Calibration error lower is better

Clairic
Cloudflare
NuExtract3
NuMind
GLiNER2
Fastino

Expected calibration error of the model's confidence: how far its stated confidence is from how often it's right.

80% range coverage should be close to 80%

Clairic
Cloudflare
NuExtract3
NuMind
GLiNER2
Fastino

Share of true values that land inside the model's 80% range. Close to 80% is good; far above or below is not.

Test items and precision

Items
Every scored field for calibration; 200 estimate items per test set for the 80% ranges
Precision
Calibration is precise. The 80% range check is about ±5.5 points per test set and ±4 pooled, enough to tell 70% from 80% from 90% only when both sets are pooled.

Full method and scoring rules