benchgap
Calibration

DeepSWE → SWE-bench Pro

SWE-bench Pro is estimated from DeepSWE with a Hill curve fitted on 20 models measured on both: y = 0.5852 + (1.2000 − 0.5852)·x^6.00 / (0.85084^6.00 + x^6.00), R² = 0.43, cross-validated error 7.8 pp. It is used for 2 estimates.

Estimated modelDeepSWESWE-bench ProSource
Pareto 26.10 Preview69.9%73.0%estimated ± 7.8 pp, low confidence
Pareto 26.974.0%77.1%estimated ± 7.8 pp, low confidence