docs/scoring.md

Scoring

DocuBench scores schema-guided extraction output against hand-verified JSON labels. The scorer is intentionally small and inspectable; the canonical implementation is scorer.py.

Inputs

Each scoring call takes:

• result: the system output JSON object, usually from the top-level data key in results/<system>/<doc_id>.json
• schema: the JSON Schema in schemas/<doc_id>.json
• label: the hand-verified target JSON in labels/<doc_id>.json

The schema determines which fields are scored.

Normalization

Before comparison:

• numeric values are cast to floats and rounded to six decimal places
• strings are lowercased
• whitespace is removed from strings
• punctuation and non-word separators are removed from strings
• empty strings, empty arrays, and empty objects are stripped
• null is preserved because it can represent an intentional labeled value

This normalization is meant to avoid penalizing harmless formatting differences. It is not semantic matching.

Non-Array Fields

Non-array leaves are scored as binary exact matches after normalization.

• matching leaf: 1
• mismatching leaf: 0
• both sides empty: skipped
• one side empty: mismatch

The non-array score is correct / total over scored leaf fields.

Array Fields

Arrays are scored order-independently.

1. Each labeled item is compared with each predicted item.
2. Item similarity is computed as binary leaf-field accuracy over the union of leaves present in either item.
3. The scorer greedily assigns the highest-scoring label/result item pairs without reusing either side.
4. Unmatched label or result items count against the score through the denominator.

Array scores are weighted by the average non-empty leaf count in labeled array items. This keeps the final score approximately leaf-weighted rather than array-count-weighted.

Aggregate Metric

For one document, the final score is a weighted average across array and non-array components.

For a system, the headline benchmark score is the macro average of per-document final scores. Each document contributes equally to the headline aggregate.

Known Tradeoffs

• Array matching is greedy, not globally optimal.
• String matching is normalized exact matching, not semantic equivalence.
• Extra fields outside the schema are ignored for non-array objects.
• Extra fields inside array items can affect item-level similarity because array item comparison uses the union of leaves.
• Field-specific tolerances are not yet expressed in schema metadata.

These tradeoffs are documented so scores are interpretable and reproducible. Changes to scoring behavior should be treated as benchmark-version changes.