OCR on DocVQA, ChartMuseum, and OCRBenchV2 (held-out)
50.05AccuracySKILLRATER
Evaluation results
| Method | Links | |
|---|---|---|
| SKILLRATERTraining Steps=100k, Model Scale=2B, Training Recipe=Isaac 0.2, Rater Parameters=150M, Rater Configuration=capability-specific2026.02 | 50.05 | |
| DataRaterTraining Steps=100k, Model Scale=2B, Training Recipe=Isaac 0.2, Rater Parameters=150M, Rater Configuration=monolithic2026.02 | 48.47 | |
| Mammoth (Baseline)Training Steps=100k, Model Scale=2B, Training Recipe=Isaac 0.2, Training Dataset=Mammoth2026.02 | 48.05 | |
| SKILLRATER CombinedTraining Steps=100k, Model Scale=2B, Training Recipe=Isaac 0.2, Rater Parameters=150M, Rater Configuration=pooled-validation2026.02 | 47.28 |