Visual Question Answering on OCR-VQA (Exact match (EM))
77.8Exact Match (EM)PaLI-3
Evaluation Results
| Method | Links | |
|---|---|---|
| PaLI-3OCR pipeline input=true, Resolution=812x8122023.10 | 77.8 | |
| PaLI-XOCR pipeline input=true, Resolution=812x8122023.10 | 77.3 | |
| PaLI-3OCR pipeline input=false, Resolution=812x8122023.10 | 76.7 | |
| DATAPROPHETRL Post-training Setup=Real RL, Data Selection Strategy=D.P., Backbone=Qwen2.5VL-3B-Instruct2026.03 | 76 | |
| PaLI-XOCR pipeline input=false, Resolution=812x8122023.10 | 75 | |
| BaselineRL Post-training Setup=Real RL, Data Selection Strategy=Equal, Backbone=Qwen2.5VL-3B-Instruct2026.03 | 75 | |
| DATAPROPHETRL Post-training Setup=Synthetic RL, Data Selection Strategy=D.P., Backbone=Qwen2.5VL-3B-Instruct2026.03 | 73 | |
| BaselineRL Post-training Setup=No RL, Data Selection Strategy=Baseline, Backbone=Qwen2.5VL-3B-Instruct2026.03 | 72 | |
| BaselineRL Post-training Setup=Synthetic RL, Data Selection Strategy=Random, Backbone=Qwen2.5VL-3B-Instruct2026.03 | 71 |