Calibration
RealWorldQA → ERQA
ERQA is estimated from RealWorldQA with a Hill curve fitted on 6 models measured on both: y = 0.0000 + (1.2000 − 0.0000)·x^6.00 / (0.82186^6.00 + x^6.00), R² = 0.71, cross-validated error 3.2 pp. It is used for 6 estimates.
| Estimated model | RealWorldQA | ERQA | Source |
|---|---|---|---|
| LFM2.5-VL-3B | 73.1% | 39.7% | estimated ± 3.2 pp, low confidence |
| LFM2.5-VL-450M | 58.4% | 13.7% | estimated ± 3.2 pp, low confidence |
| North Micro Vision Instruct | 62.2% | 19.0% | estimated ± 3.2 pp, low confidence |
| Qwen3.6-35B-A3B | 85.3% | 66.7% | estimated ± 3.2 pp, medium confidence |
| Ternary Bonsai 2 27B | 80.1% | 55.4% | estimated ± 3.2 pp, low confidence |
| ZAYA1-VL-8B | 65.0% | 23.6% | estimated ± 3.2 pp, low confidence |