benchgap
Calibration

SciCode → SWE-bench Verified

SWE-bench Verified is estimated from SciCode with a linear curve fitted on 14 models measured on both: y = 1.4035·x + 0.0883, R² = 0.83, cross-validated error 5.7 pp. It is used for 4 estimates.

Estimated modelSciCodeSWE-bench VerifiedSource
A.X K241.0%66.4%estimated ± 5.7 pp, medium confidence
Agents-A1-4B29.6%50.4%estimated ± 5.7 pp, medium confidence
Granite 4.2 3B24.1%42.7%estimated ± 5.7 pp, low confidence
Mercury 2.538.0%62.2%estimated ± 5.7 pp, medium confidence