GLAM extraction benchmark
Overall extraction quality
2 datasets · 10 models with complete results
Overall F1 is the average of the dataset F1 scores. Each dataset has equal weight (50% each). Only models evaluated on every selected dataset are included.
Choose a dataset above to inspect its results and methodology. These are small evaluation sets; nearby scores should not be read as a definitive ranking.
Uses total parameters, including for mixture-of-experts models. Click a column heading to sort; click again to reverse.
| GLM-5.3-Flash 321.32B | 91.5 |
| Qwen3.8-27B 27.78B | 88.8 |
| Qwen3-VL-235B 235.67B | 87.9 |
| Qwen3.5-4B 4.660B | 84.5 |
| NuExtract-3 4.539B | 83.7 |
| Qwen3.5-9B 9.65B | 83.6 |
| Qwen3-VL-8B 8.767B | 83.3 |
| Granite-Vision-4.1-4B 3.997B | 76.8 |
| gemma-4-E4B 7.996B | 72.9 |
| Qwen3.5-2B 2.274B | 63.3 |
No models match this size limit.
Scores are shown on a 0–100 scale. Dataset weights, component scores and pinned revisions are recorded in the score download.
Size vs. overall extraction quality
Upper-left is better: smaller models, higher overall F1. Teal points show the observed Pareto frontier.
Sources and reproducibility
Overall scores and component provenance · Dataset configurations and revisions · Code · Raw predictions
The overall view combines the selected dataset results; it does not create a new dataset or change individual scores.