Calibration
AA-SciCode → PostTrainBench v1.1
PostTrainBench v1.1 is estimated from AA-SciCode with a Hill curve fitted on 12 models measured on both: y = 0.2136 + (1.2000 − 0.2136)·x^6.00 / (0.79405^6.00 + x^6.00), R² = 0.41, cross-validated error 7.2 pp. It is used for 12 estimates.
| Estimated model | AA-SciCode | PostTrainBench v1.1 | Source |
|---|---|---|---|
| A.X K2 | 41.0% | 23.2% | estimated ± 7.2 pp, low confidence |
| DeepSeek V3 0324 | 39.0% | 22.7% | estimated ± 7.2 pp, low confidence |
| Granite 4.2 3B | 25.3% | 21.5% | estimated ± 7.2 pp, low confidence |
| K-EXAONE 2.0 | 42.0% | 23.5% | estimated ± 7.2 pp, low confidence |
| Ling 3.0 Flash VL | 44.2% | 24.2% | estimated ± 7.2 pp, low confidence |
| Ling 3.0 Tiny | 24.2% | 21.4% | estimated ± 7.2 pp, low confidence |
| Ling 3.1 Flash | 54.1% | 30.3% | estimated ± 7.2 pp, low confidence |
| Mercury 2.5 | 38.5% | 22.6% | estimated ± 7.2 pp, low confidence |
| Mistral Large 4 | 54.2% | 30.4% | estimated ± 7.2 pp, low confidence |
| North Mini Code | 38.8% | 22.7% | estimated ± 7.2 pp, low confidence |
| Solar Pro 3 | 25.5% | 21.5% | estimated ± 7.2 pp, low confidence |
| Solar Pro 4 | 44.6% | 24.4% | estimated ± 7.2 pp, low confidence |