benchgap
Calibration

BFCL v4 → τ²-bench results

τ²-bench results is estimated from BFCL v4 with a Michaelis–Menten curve fitted on 5 models measured on both: y = 2.0000·x / (0.85488 + x), R² = 0.54, cross-validated error 14.1 pp. It is used for 15 estimates.

Estimated modelBFCL v4τ²-bench resultsSource
Atria Dawn Preview77.0%94.8%estimated ± 14.1 pp, low confidence
BTL-388.5%100.0%estimated ± 14.1 pp, low confidence
BTL-473.5%92.5%estimated ± 14.1 pp, low confidence
Granite 4.2 30B61.4%83.6%estimated ± 14.1 pp, low confidence
Granite 4.2 3B52.4%76.0%estimated ± 14.1 pp, low confidence
LFM2.5-2.6B56.9%79.9%estimated ± 14.1 pp, low confidence
LFM2.5-230M21.0%39.5%estimated ± 14.1 pp, low confidence
LFM2.5-VL-3B32.5%55.1%estimated ± 14.1 pp, low confidence
LFM2.5-VL-450M21.1%39.6%estimated ± 14.1 pp, low confidence
Mellum2-12B-A2.5B-Instruct44.2%68.2%estimated ± 14.1 pp, low confidence
Mellum2-12B-A2.5B-Thinking45.6%69.6%estimated ± 14.1 pp, low confidence
MiniCPM5-1B25.2%45.5%estimated ± 14.1 pp, low confidence
MiniCPM5-2B66.6%87.6%estimated ± 14.1 pp, low confidence
Pokee-Isaac 28B70.9%90.7%estimated ± 14.1 pp, low confidence
ZAYA1-8B39.2%62.9%estimated ± 14.1 pp, low confidence