| # ParseBench reproduction | |
| See [pipeline/README.md](pipeline/README.md) for the complete runnable bundle: | |
| - Prepare the public ParseBench dataset with the evaluated rendering settings. | |
| - Serve the pinned Nemotron 3.5 Parse checkpoint on one NVIDIA Hopper H100. | |
| - Run single-, double-, or triple-pass inference with the matching postprocessing. | |
| - Score final predictions using the pinned current official ParseBench implementation. | |
| | Pipeline | Reproduced ParseBench score | | |
| | --- | ---: | | |
| | Single | 72.72 | | |
| | Double | 74.88 | | |
| | Triple | 75.14 | | |
| | Triple, region union | 75.20 | | |
| [Expected results](pipeline/expected_results.json) record exact identities and settings. [Validation metrics](pipeline/validation_metrics.json) record full-run coverage, category scores, and the distinction between full generation, postprocessing replay, fresh official scoring, and GPU smoke testing. | |
| Historical artifacts elsewhere in this directory retain their own evaluation contracts. | |
| The headline score uses triple-pass inference with `--triple-regions union`. Default triple-pass inference uses detector regions only; region union adds full-page Table/Chart regions subject to overlap filtering. | |