58-model evaluation · three matched tracks
PureDocBench Leaderboard
Document parsing, measured across clean, digitally degraded,
and real-degraded documents.
58Current models
3Image tracks
1,475Base pages
Loading results…Click a column to sort
| Loading the leaderboard results… |
↑ Higher is better · ↓ Lower is better · All downloads include the complete benchmark snapshot.
How to read the scores
Overall = [100 × (1 − TextEdit) + FormulaCDM + TableTEDS] / 3. Reading-order edit distance is reported separately. Avg3 averages the three track Overall scores. TextEdit and reading-order edit distance use a 0–1 scale; the other scores use a 0–100 scale.
Source & scope
The default view shows the current 58-model release. A historical community submission is available separately; its badge links to the source. Each track contains 1,475 images. Scores retain their supplied evaluation version and precision. Full result tables ↗