Datasets:
Limitations
DocuBench is intentionally hard, but it is not exhaustive.
Dataset Size
The benchmark has 50 documents. That is enough for careful inspection and regression testing, but not enough to support broad statistical claims about every document domain.
Public-Document Bias
All documents are public, publicly posted samples, openly licensed files, government publications, or benchmark-authored artifacts. This improves reproducibility, but it means the corpus may differ from private enterprise document distributions.
Label And Schema Scope
Each schema asks for selected fields rather than every possible fact in a document. A system may extract useful information that is not measured by a given schema.
Scoring Scope
The current scorer uses normalized exact matching and greedy array alignment. It does not yet support field-specific semantic matching, numeric tolerances beyond float normalization, or globally optimal array assignment.
Aggregate Interpretation
The headline score is a macro average over documents. Users should inspect per-document and per-capability results before making system decisions.
Cost And Latency
Some result files include cost and timing metadata, but the benchmark does not yet enforce a uniform cost/latency reporting contract across all systems.