Calibration
MMLU-Redux → SuperGPQA
SuperGPQA is estimated from MMLU-Redux with a Hill curve fitted on 8 models measured on both: y = 0.0000 + (1.2000 − 0.0000)·x^6.00 / (0.89988^6.00 + x^6.00), R² = 0.96, cross-validated error 4.5 pp. It is used for 3 estimates.
| Estimated model | MMLU-Redux | SuperGPQA | Source |
|---|---|---|---|
| Mellum2-12B-A2.5B-Instruct | 78.1% | 35.9% | estimated ± 4.5 pp, high confidence |
| Mellum2-12B-A2.5B-Thinking | 86.2% | 52.3% | estimated ± 4.5 pp, high confidence |
| Ternary Bonsai 2 27B | 89.1% | 58.2% | estimated ± 4.5 pp, high confidence |