Benchmark — Tablewise vs MarkItDown

Scoreboard

Each cell is Tablewise / MarkItDown. “Unnamed” and “NaN” are defects: lower is better. Formatted percentages, marked formulas and ISO dates are wins: higher is better.

Case	Unnamed cols	NaN cells	Formatted %	Formulas marked	Hidden labelled	ISO dates
`financial_pack` Three-level header, merged cells, formulas, percentages	0 / 11	0 / 41	3 / 0	3 / 0	0 / 0	0 / 0
`invoices_dates` Dates, datetimes, percentages, decimals	0 / 0	0 / 0	4 / 0	0 / 0	0 / 0	8 / 8
`hidden_sheet` A visible report plus a hidden helper sheet	0 / 0	0 / 0	0 / 0	0 / 0	1 / 0	0 / 0
`formula_model` Formulas with no cached value	0 / 0	0 / 6	0 / 0	6 / 0	0 / 0	0 / 0
`ledger_negatives` Negative zero, tiny values, zero formatting	0 / 0	0 / 0	0 / 0	0 / 0	0 / 0	0 / 0

The same file, both ways

financial_pack: three-level header, merged cells, formulas, percentages.

MarkItDownflattened

## P&L
| Property | 2026 | Unnamed: 2 | Unnamed: 3 | Unnamed: 4 | Unnamed: 5 | Unnamed: 6 | Unnamed: 7 | Unnamed: 8 | Unnamed: 9 | Unnamed: 10 | Unnamed: 11 | Unnamed: 12 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| NaN | Actual | NaN | NaN | NaN | Budget | NaN | NaN | NaN | Last year | NaN | NaN | NaN |
| NaN | Jan | Feb | Mar | Apr | Jan | Feb | Mar | Apr | Jan | Feb | Mar | Apr |
| ADR | 120.5 | NaN | NaN | NaN | 110 | NaN | NaN | NaN | 105.25 | NaN | NaN | NaN |
| Occupancy | 0.8769 | NaN | NaN | NaN | 0.82 | NaN | NaN | NaN | 0.79 | NaN | NaN | NaN |
| RevPAR | NaN | NaN | NaN | NaN | NaN | NaN | NaN | NaN | NaN | NaN | NaN | NaN |

Tablewisereadable

## P&L

| Property | 2026 |  |  |  |  |  |  |  |  |  |  |  |
|  | Actual | Actual | Actual | Actual | Budget | Budget | Budget | Budget | Last year | Last year | Last year | Last year |
|  | Jan | Feb | Mar | Apr | Jan | Feb | Mar | Apr | Jan | Feb | Mar | Apr |
| ADR | 120.50 |  |  |  | 110 |  |  |  | 105.25 |  |  |  |
| Occupancy | 87.69% |  |  |  | 82.00% |  |  |  | 79.00% |  |  |  |
| RevPAR | (formula not calculated) |  |  |  | (formula not calculated) |  |  |  | (formula not calculated) |  |  |  |

What this does and does not prove

It proves that on these five structures, Tablewise preserves header labels, percentage formats, formula state and dates where MarkItDown does not.
It does not prove anything about PDFs or scanned documents: those go through Docling, a different engine, and are outside this benchmark.
MarkItDown is a general-purpose converter and is excellent at many things. Spreadsheet structure is not one of them, and that is the case we exist for.
Reproduce: .venv/bin/python3 bench/run_benchmark.py.