extractingbench

Math

Can it compute values that aren't printed?

One step: a sum, a date difference, a count. Multi-step: subtotal minus discount plus tax, a per-person share, a due date from terms.

Cloudflare says multi-step arithmetic is a known weakness of Clairic. This tests it. Models that can't compute show not supported.

Math, one step one operation on printed values

Clairic
Cloudflare
NuExtract3
NuMind
GLiNER2
Fastino

Accuracy on values that need one operation, such as a sum or a date difference.

Math, multi-step two or more operations

Clairic
Cloudflare
NuExtract3
NuMind
GLiNER2
Fastino

Accuracy on values that need two or more operations.

Test items and precision

Items
About 450 one-step and 450 multi-step computed fields per test set
Precision
About ±4 to ±5 points per bucket.

Full method and scoring rules