HighlightBench

HighlightBench is a diagnostic benchmark for markup-driven table understanding that decomposes evaluation into five task families. It covers table QA over visual emphasis such as highlights, underlines, bold text, color annotations, missing entries, and structured table comparisons.

Files

• highlightbench/real/qa.jsonl: real-table QA annotations
• highlightbench/real/images/: real-table images
• highlightbench/synthetic/qa.jsonl: synthetic-table QA annotations
• highlightbench/synthetic/images/: synthetic-table images
• highlightbench/highlightbench_all_qa.jsonl: combined QA annotations
• highlightbench/score_qa.py: scoring script
• highlight_generator/: minimal synthetic image generator
• review.html: local visual review page

Each QA row keeps the compact fields needed for use, scoring, and subtask-level analysis:

{"dataset":"real","qid":"...","image_path":"...","task":"Constrained Retrieval","subtask":"Cell Retrieval","question":"...","answer":"..."}

Scoring

Prediction files only need qid and answer; task and subtask are metadata for analysis.

cd highlightbench
python3 score_qa.py --gt real/qa.jsonl --pred examples/pred_sample.real.jsonl --out examples/pred_sample.real.score.json
python3 score_qa.py --gt synthetic/qa.jsonl --pred examples/pred_sample.synthetic.jsonl --out examples/pred_sample.synthetic.score.json