Calibration
BFCL v4 → GDPval-AA
GDPval-AA is estimated from BFCL v4 with a inverse Michaelis–Menten curve fitted on 8 models measured on both: y = 0.05373·(x − 0.5030) / (0.5030 + 0.2899 − x), R² = 0.94, cross-validated error 5.8 pp. It is used for 13 estimates.
| Estimated model | BFCL v4 | GDPval-AA | Source |
|---|---|---|---|
| BTL-3 | 88.5% | 100.0% | estimated ± 5.8 pp, low confidence |
| BTL-4 | 73.5% | 21.5% | estimated ± 5.8 pp, medium confidence |
| LFM2.5-230M | 21.0% | 0.0% | estimated ± 5.8 pp, low confidence |
| LFM2.5-8B-A1B | 49.7% | 0.0% | estimated ± 5.8 pp, low confidence |
| LFM2.5-VL-3B | 32.5% | 0.0% | estimated ± 5.8 pp, low confidence |
| LFM2.5-VL-450M | 21.1% | 0.0% | estimated ± 5.8 pp, low confidence |
| LLaDA2.2-flash | 60.8% | 3.0% | estimated ± 5.8 pp, medium confidence |
| LLaDA2.2-mini | 47.7% | 0.0% | estimated ± 5.8 pp, low confidence |
| Mellum2-12B-A2.5B-Instruct | 44.2% | 0.0% | estimated ± 5.8 pp, low confidence |
| Mellum2-12B-A2.5B-Thinking | 45.6% | 0.0% | estimated ± 5.8 pp, low confidence |
| MiniCPM5-1B | 25.2% | 0.0% | estimated ± 5.8 pp, low confidence |
| Pokee-Isaac 28B | 70.9% | 13.3% | estimated ± 5.8 pp, medium confidence |
| ZAYA1-8B | 39.2% | 0.0% | estimated ± 5.8 pp, low confidence |