benchgap
Calibration

IFBench → IFEval

IFEval is estimated from IFBench with a offset logistic curve fitted on 16 models measured on both: y = 0.7828 + (0.9282 − 0.7828) / (1 + exp(−40.88·(x − 0.5221))), R² = 0.84, cross-validated error 3.5 pp. It is used for 21 estimates.

Estimated modelIFBenchIFEvalSource
A.X K275.9%92.8%estimated ± 3.5 pp, high confidence
Granite 4.2 30B77.2%92.8%estimated ± 3.5 pp, high confidence
Granite 4.2 3B74.3%92.8%estimated ± 3.5 pp, high confidence
Granite 4.2 8B79.3%92.8%estimated ± 3.5 pp, high confidence
Hy3 Preview63.1%92.7%estimated ± 3.5 pp, high confidence
Inkling79.8%92.8%estimated ± 3.5 pp, high confidence
Inkling-Small82.2%92.8%estimated ± 3.5 pp, medium confidence
LFM2.5-2.6B59.2%92.0%estimated ± 3.5 pp, high confidence
Ling 3.0 Flash74.5%92.8%estimated ± 3.5 pp, high confidence
Ling 3.0 Flash FP873.4%92.8%estimated ± 3.5 pp, high confidence
LLaDA2.2-mini24.9%78.3%estimated ± 3.5 pp, medium confidence
MAI-Thinking-185.0%92.8%estimated ± 3.5 pp, medium confidence
Mercury 2.577.0%92.8%estimated ± 3.5 pp, high confidence
Muse Glimmer 30B77.0%92.8%estimated ± 3.5 pp, high confidence
Nemotron 3.5 Lightning 30B A3B NVFP472.9%92.8%estimated ± 3.5 pp, high confidence
Qwen3.8-27B79.5%92.8%estimated ± 3.5 pp, high confidence
Qwen3.8-Flash-Next81.3%92.8%estimated ± 3.5 pp, medium confidence
Qwen3.8 Max82.8%92.8%estimated ± 3.5 pp, medium confidence
Qwen3.8-Omni-Flash81.5%92.8%estimated ± 3.5 pp, medium confidence
Beam79.7%92.8%estimated ± 3.5 pp, high confidence
Solar Open 280.0%92.8%estimated ± 3.5 pp, high confidence