Calibration
LiveCodeBench v6 → SciCode
SciCode is estimated from LiveCodeBench v6 with a inverse Michaelis–Menten curve fitted on 11 models measured on both: y = 0.34217·(x − 0.0000) / (0.0000 + 1.4868 − x), R² = 0.88, cross-validated error 5.4 pp. It is used for 6 estimates.
| Estimated model | LiveCodeBench v6 | SciCode | Source |
|---|---|---|---|
| BTL-3 | 88.1% | 49.8% | estimated ± 5.4 pp, medium confidence |
| LLaDA2.2-mini | 28.1% | 8.0% | estimated ± 5.4 pp, low confidence |
| Mellum2-12B-A2.5B-Instruct | 37.2% | 11.4% | estimated ± 5.4 pp, low confidence |
| Mellum2-12B-A2.5B-Thinking | 69.9% | 30.4% | estimated ± 5.4 pp, medium confidence |
| MiniCPM5-1B | 33.5% | 10.0% | estimated ± 5.4 pp, low confidence |
| ZAYA1-8B | 65.8% | 27.2% | estimated ± 5.4 pp, medium confidence |