Image-only 'scanned' documents paired with their text source, for measuring OCR accuracy against a known ground truth.
A clean bar chart of monthly revenue with axis labels, gridlines, and value labels: a real-world fixture for testing chart-extraction and OCR tools, vision models, and thumbnail rendering.
A two-series line chart (desktop vs. mobile visitors) with markers, a legend, and gridlines, for testing chart parsing, trend extraction, and image pipelines against a known plot.
A five-slice pie chart of traffic sources with percentage labels and a legend: a real-world fixture for chart-recognition, segmentation, and OCR of embedded text.
A rendered business card with a name, title, company, email, phone, and address, paired with a matching vCard so you can measure OCR and contact-extraction accuracy against a known ground truth.
The vCard (.vcf) ground truth for the business-card image: the same name, title, company, email, phone, and address in structured form. The positive reference for testing OCR-to-contact extraction.
Rendered SAMPLE data table as a PNG, for table-detection / OCR vision pipelines.
Rendered SAMPLE data table as a PNG, for table-detection / OCR vision pipelines.
Rendered SAMPLE data table as a PNG, for table-detection / OCR vision pipelines.
Synthetic dashboard mock (KPI cards + bar chart): SAMPLE UI for layout / chart vision models. Not a real product.
Synthetic dashboard mock (KPI cards + bar chart): SAMPLE UI for layout / chart vision models. Not a real product.
Synthetic dashboard mock (KPI cards + bar chart): SAMPLE UI for layout / chart vision models. Not a real product.
Synthetic dashboard mock (KPI cards + bar chart): SAMPLE UI for layout / chart vision models. Not a real product.
Synthetic dashboard mock (KPI cards + bar chart): SAMPLE UI for layout / chart vision models. Not a real product.
Synthetic dashboard mock (KPI cards + bar chart): SAMPLE UI for layout / chart vision models. Not a real product.
Hard OCR case (skew): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Exact ground-truth string for the skew-invoice OCR hard case.
Hard OCR case (low-contrast): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Exact ground-truth string for the low-contrast OCR hard case.
Hard OCR case (dotted): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Exact ground-truth string for the dotted-serial OCR hard case.
Hard OCR case (skew): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Exact ground-truth string for the skew-address OCR hard case.
Hard OCR case (low-contrast): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Exact ground-truth string for the low-receipt OCR hard case.
Hard OCR case (dotted): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Exact ground-truth string for the dotted-code OCR hard case.
Hard OCR case (skew): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Exact ground-truth string for the skew-date OCR hard case.
Hard OCR case (low-contrast): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Exact ground-truth string for the low-label OCR hard case.
A delivery typeset over a synthetic photographic background, gaussian blur. Softened by a recorded Gaussian radius, which is what an out-of-focus capture looks like. 29 words, each with an exact bounding box in the companion JSON.
A delivery typeset over a synthetic photographic background, undistorted. Straight-on and undamaged, the baseline every other variant in this group is measured against. 29 words, each with an exact bounding box in the companion JSON.
A delivery typeset over a synthetic photographic background, low contrast. Gain and offset applied so the text sits close to the ground, which is where thresholding fails. 29 words, each with an exact bounding box in the companion JSON.
A delivery typeset over a synthetic photographic background, sensor noise. Additive noise at a recorded sigma, as a dim-light capture would carry. 29 words, each with an exact bounding box in the companion JSON.
A delivery typeset over a synthetic photographic background, rotated +3 degrees. A slight skew of the kind a phone photo of a docket actually has. 29 words, each with an exact bounding box in the companion JSON.
A delivery typeset over a synthetic photographic background, rotated -7 degrees. Rotated, with every word box rotated with it, so skew correction can be scored rather than eyeballed. 29 words, each with an exact bounding box in the companion JSON.
A invoice typeset over a synthetic photographic background, gaussian blur. Softened by a recorded Gaussian radius, which is what an out-of-focus capture looks like. 33 words, each with an exact bounding box in the companion JSON.
A invoice typeset over a synthetic photographic background, undistorted. Straight-on and undamaged, the baseline every other variant in this group is measured against. 33 words, each with an exact bounding box in the companion JSON.
A invoice typeset over a synthetic photographic background, low contrast. Gain and offset applied so the text sits close to the ground, which is where thresholding fails. 33 words, each with an exact bounding box in the companion JSON.
A invoice typeset over a synthetic photographic background, sensor noise. Additive noise at a recorded sigma, as a dim-light capture would carry. 33 words, each with an exact bounding box in the companion JSON.
A invoice typeset over a synthetic photographic background, rotated +3 degrees. A slight skew of the kind a phone photo of a docket actually has. 33 words, each with an exact bounding box in the companion JSON.
A invoice typeset over a synthetic photographic background, rotated -7 degrees. Rotated, with every word box rotated with it, so skew correction can be scored rather than eyeballed. 33 words, each with an exact bounding box in the companion JSON.
A receipt typeset over a synthetic photographic background, gaussian blur. Softened by a recorded Gaussian radius, which is what an out-of-focus capture looks like. 23 words, each with an exact bounding box in the companion JSON.
A receipt typeset over a synthetic photographic background, undistorted. Straight-on and undamaged, the baseline every other variant in this group is measured against. 23 words, each with an exact bounding box in the companion JSON.
A receipt typeset over a synthetic photographic background, low contrast. Gain and offset applied so the text sits close to the ground, which is where thresholding fails. 23 words, each with an exact bounding box in the companion JSON.
A receipt typeset over a synthetic photographic background, sensor noise. Additive noise at a recorded sigma, as a dim-light capture would carry. 23 words, each with an exact bounding box in the companion JSON.
A receipt typeset over a synthetic photographic background, rotated +3 degrees. A slight skew of the kind a phone photo of a docket actually has. 23 words, each with an exact bounding box in the companion JSON.
A receipt typeset over a synthetic photographic background, rotated -7 degrees. Rotated, with every word box rotated with it, so skew correction can be scored rather than eyeballed. 23 words, each with an exact bounding box in the companion JSON.
An image-only PDF containing a rasterised 'scan' of the simple document, with no text layer. Paired with the text version so you can score OCR output against a known ground truth.
A realistic café receipt as a searchable PDF with a real text layer: the OCR ground truth paired with an image-only scanned twin, so you can score OCR output against known text.
The same café receipt as an image-only 'scan': no text layer, with a slight rotation and grain so it reads like a photographed receipt. Paired with the searchable version as OCR ground truth.
An airline boarding pass laid out as a landscape PDF: passenger, flight, gate, seat, PNR, and a decorative barcode strip. A fixture for testing document-layout parsing and field extraction. Synthetic; the barcode is decorative.
A searchable PDF invoice with selectable text: the OCR ground-truth twin of the scanned JPEG. Fictional SAMPLE content for measuring OCR accuracy.
A scanned-look JPEG of a invoice. Pair with the searchable PDF twin to score OCR against known text. Synthetic SAMPLE content only.
A searchable PDF utility bill with selectable text: the OCR ground-truth twin of the scanned JPEG. Fictional SAMPLE content for measuring OCR accuracy.
A scanned-look JPEG of a utility bill. Pair with the searchable PDF twin to score OCR against known text. Synthetic SAMPLE content only.
A searchable PDF id card mock with selectable text: the OCR ground-truth twin of the scanned JPEG. Fictional SAMPLE content for measuring OCR accuracy.
A scanned-look JPEG of a id card mock. Pair with the searchable PDF twin to score OCR against known text. Synthetic SAMPLE content only.
A searchable PDF multi-column article with selectable text: the OCR ground-truth twin of the scanned JPEG. Fictional SAMPLE content for measuring OCR accuracy.
A scanned-look JPEG of a multi-column article. Pair with the searchable PDF twin to score OCR against known text. Synthetic SAMPLE content only.
A searchable PDF handwritten-print form with selectable text: the OCR ground-truth twin of the scanned JPEG. Fictional SAMPLE content for measuring OCR accuracy.
A scanned-look JPEG of a handwritten-print form. Pair with the searchable PDF twin to score OCR against known text. Synthetic SAMPLE content only.
A searchable PDF pos receipt with selectable text: the OCR ground-truth twin of the scanned JPEG. Fictional SAMPLE content for measuring OCR accuracy.
A scanned-look JPEG of a pos receipt. Pair with the searchable PDF twin to score OCR against known text. Synthetic SAMPLE content only.
A single-line latin script image for OCR engines. Pair with the UTF-8 ground-truth text twin.
UTF-8 ground-truth transcript for the latin OCR line image: exact expected OCR output.
A single-line cjk script image for OCR engines. Pair with the UTF-8 ground-truth text twin.
UTF-8 ground-truth transcript for the cjk OCR line image: exact expected OCR output.
A single-line arabic script image for OCR engines. Pair with the UTF-8 ground-truth text twin.
UTF-8 ground-truth transcript for the arabic OCR line image: exact expected OCR output.
A PDF with clearly labelled fictional SAMPLE PII fields: the unredacted twin for testing redaction tooling.
JPEG twin of the unredacted SAMPLE PII page, for image-based redaction and OCR pipelines.
JPEG twin with SAMPLE PII fields blacked out. Pair with the clear JPEG to score redaction tools.
Costing report PDF for Alder Table Bistro. A 3-page report whose first page begins "Recipe and plate costing". Page 1 lists the 6 purchase packs with their converted unit costs and opens the per-dish costing, which runs through page 2 for all 4 menu items. Page 3 carries the period cost of goods sold (1574.81 CAD) and the 5 yield tests.
Costing report PDF for Juniper Corner Cafe. A 3-page report whose first page begins "Recipe and plate costing". Page 1 lists the 6 purchase packs with their converted unit costs and opens the per-dish costing, which runs through page 2 for all 4 menu items. Page 3 carries the period cost of goods sold (1538.14 CAD) and the 2 yield tests.
Costing report PDF for Copper Oven Bakery. A 3-page report whose first page begins "Recipe and plate costing". Page 1 lists the 6 purchase packs with their converted unit costs and opens the per-dish costing, which runs through page 2 for all 4 menu items. Page 3 carries the period cost of goods sold (623.44 CAD) and the 2 yield tests.
Costing report PDF for Ember Lane Pizza. A 3-page report whose first page begins "Recipe and plate costing". Page 1 lists the 6 purchase packs with their converted unit costs and opens the per-dish costing, which runs through page 2 for all 4 menu items. Page 3 carries the period cost of goods sold (3110.43 CAD) and the 2 yield tests.
Costing report PDF for Cedar Street Tacos. A 3-page report whose first page begins "Recipe and plate costing". Page 1 lists the 6 purchase packs with their converted unit costs and opens the per-dish costing, which runs through page 2 for all 4 menu items. Page 3 carries the period cost of goods sold (2559.57 CAD) and the 3 yield tests.
Costing report PDF for Willow Event Kitchen. A 3-page report whose first page begins "Recipe and plate costing". Page 1 lists the 6 purchase packs with their converted unit costs and opens the per-dish costing, which runs through page 2 for all 4 menu items. Page 3 carries the period cost of goods sold (2106.63 CAD) and the 2 yield tests.
Costing report PDF for Harbour Kitchen Group. A 3-page report whose first page begins "Recipe and plate costing". Page 1 lists the 6 purchase packs with their converted unit costs and opens the per-dish costing, which runs through page 2 for all 4 menu items. Page 3 carries the period cost of goods sold (3241.08 CAD) and the 4 yield tests.
Printed menu for Alder Table Bistro. A one page food menu with 3 sections (Starters, Mains, Sides) and 4 dishes, prices right aligned in CAD and an allergen number after every dish. First page text begins "Alder Table Bistro". Every price is the menu_price_cad in plate-costs.csv and every allergen number is a Y column in allergen-matrix.csv.
Allergen menu for Alder Table Bistro. A landscape page carrying 4 dishes against the 14 declarable allergens, Y for declared and a hyphen for not declared, with a second table naming the ingredient behind every Y. The grid reproduces allergen-matrix.csv in the costing folder cell for cell.
Drinks list for Alder Table Bistro. A one page list with 4 lines by the glass and 4 by the bottle. Where a drink is sold both ways the glass price is the bottle price divided by 5 pours and rounded up to the next 0.25, which the page states. Drinks sit outside the shared ingredient costing study, so no plate cost is claimed for any of them.
Daily specials insert for Alder Table Bistro. A dated single page for service on 2026-08-28, inside the costing period 2026-08-20 to 2026-09-02. 2 dishes from the printed menu at 85 percent of the menu price rounded to the nearest 0.05, with the menu price and the saving shown beside each.
Set lunch for Alder Table Bistro. A fixed price page at 31.00 CAD covering 2 dishes that are already on the printed menu. The same items ordered separately come to 35.50 CAD, so the page states a saving of 4.50 CAD. Plate cost for the bundle is 4.33 CAD.
Children's plates for Alder Table Bistro. A one page children's menu: the same 4 dishes as smaller portions at 60 percent of the menu price, rounded to the nearest 0.05. The prices are 14.40, 6.90, 7.80, 7.50 CAD, and every one already lands on a five cent step, so no row on the page is rounded.
Printed menu for Juniper Corner Cafe. A one page food menu with 2 sections (Espresso bar, All day kitchen) and 4 dishes, prices right aligned in CAD and an allergen number after every dish. First page text begins "Juniper Corner Cafe". Every price is the menu_price_cad in plate-costs.csv and every allergen number is a Y column in allergen-matrix.csv.
Allergen menu for Juniper Corner Cafe. A landscape page carrying 4 dishes against the 14 declarable allergens, Y for declared and a hyphen for not declared, with a second table naming the ingredient behind every Y. The grid reproduces allergen-matrix.csv in the costing folder cell for cell.
Drinks list for Juniper Corner Cafe. A one page list with 4 lines by the glass and 4 by the bottle. Where a drink is sold both ways the glass price is the bottle price divided by 5 pours and rounded up to the next 0.25, which the page states. Drinks sit outside the shared ingredient costing study, so no plate cost is claimed for any of them.
Daily specials insert for Juniper Corner Cafe. A dated single page for service on 2026-08-28, inside the costing period 2026-08-20 to 2026-09-02. 2 dishes from the printed menu at 85 percent of the menu price rounded to the nearest 0.05, with the menu price and the saving shown beside each.
Breakfast set for Juniper Corner Cafe. A fixed price page at 13.50 CAD covering 2 dishes that are already on the printed menu. The same items ordered separately come to 15.25 CAD, so the page states a saving of 1.75 CAD. Plate cost for the bundle is 2.22 CAD.
Printed menu for Copper Oven Bakery. A one page food menu with 2 sections (Bread counter, Viennoiserie and sweet) and 4 dishes, prices right aligned in CAD and an allergen number after every dish. First page text begins "Copper Oven Bakery". Every price is the menu_price_cad in plate-costs.csv and every allergen number is a Y column in allergen-matrix.csv.
Allergen menu for Copper Oven Bakery. A landscape page carrying 4 dishes against the 14 declarable allergens, Y for declared and a hyphen for not declared, with a second table naming the ingredient behind every Y. The grid reproduces allergen-matrix.csv in the costing folder cell for cell.
Drinks list for Copper Oven Bakery. A one page list with 4 lines by the glass and 4 by the bottle. Where a drink is sold both ways the glass price is the bottle price divided by 5 pours and rounded up to the next 0.25, which the page states. Drinks sit outside the shared ingredient costing study, so no plate cost is claimed for any of them.
Daily specials insert for Copper Oven Bakery. A dated single page for service on 2026-08-28, inside the costing period 2026-08-20 to 2026-09-02. 2 dishes from the printed menu at 85 percent of the menu price rounded to the nearest 0.05, with the menu price and the saving shown beside each.
Pastry box of three for Copper Oven Bakery. A fixed price page at 11.50 CAD covering 3 dishes that are already on the printed menu. The same items ordered separately come to 13.50 CAD, so the page states a saving of 2.00 CAD. Plate cost for the bundle is 1.24 CAD.
Printed menu for Ember Lane Pizza. A one page food menu with 2 sections (Classic pizzas, House pizzas) and 4 dishes, prices right aligned in CAD and an allergen number after every dish. First page text begins "Ember Lane Pizza". Every price is the menu_price_cad in plate-costs.csv and every allergen number is a Y column in allergen-matrix.csv.
Allergen menu for Ember Lane Pizza. A landscape page carrying 4 dishes against the 14 declarable allergens, Y for declared and a hyphen for not declared, with a second table naming the ingredient behind every Y. The grid reproduces allergen-matrix.csv in the costing folder cell for cell.
Drinks list for Ember Lane Pizza. A one page list with 4 lines by the glass and 4 by the bottle. Where a drink is sold both ways the glass price is the bottle price divided by 5 pours and rounded up to the next 0.25, which the page states. Drinks sit outside the shared ingredient costing study, so no plate cost is claimed for any of them.
Daily specials insert for Ember Lane Pizza. A dated single page for service on 2026-08-28, inside the costing period 2026-08-20 to 2026-09-02. 2 dishes from the printed menu at 85 percent of the menu price rounded to the nearest 0.05, with the menu price and the saving shown beside each.
Two pizza family deal for Ember Lane Pizza. A fixed price page at 33.00 CAD covering 2 dishes that are already on the printed menu. The same items ordered separately come to 37.00 CAD, so the page states a saving of 4.00 CAD. Plate cost for the bundle is 4.79 CAD.
Children's pizzas for Ember Lane Pizza. A one page children's menu: the same 4 dishes as smaller portions at 60 percent of the menu price, rounded to the nearest 0.05. The prices are 10.20, 12.00, 11.10, 10.80 CAD, and every one already lands on a five cent step, so no row on the page is rounded.
Printed menu for Cedar Street Tacos. A one page food menu with 2 sections (Tacos, Bowls) and 4 dishes, prices right aligned in CAD and an allergen number after every dish. First page text begins "Cedar Street Tacos". Every price is the menu_price_cad in plate-costs.csv and every allergen number is a Y column in allergen-matrix.csv.
Allergen menu for Cedar Street Tacos. A landscape page carrying 4 dishes against the 14 declarable allergens, Y for declared and a hyphen for not declared, with a second table naming the ingredient behind every Y. The grid reproduces allergen-matrix.csv in the costing folder cell for cell.
Drinks list for Cedar Street Tacos. A one page list with 4 lines by the glass and 4 by the bottle. Where a drink is sold both ways the glass price is the bottle price divided by 5 pours and rounded up to the next 0.25, which the page states. Drinks sit outside the shared ingredient costing study, so no plate cost is claimed for any of them.
Daily specials insert for Cedar Street Tacos. A dated single page for service on 2026-08-28, inside the costing period 2026-08-20 to 2026-09-02. 2 dishes from the printed menu at 85 percent of the menu price rounded to the nearest 0.05, with the menu price and the saving shown beside each.
Two plate combo for Cedar Street Tacos. A fixed price page at 22.50 CAD covering 2 dishes that are already on the printed menu. The same items ordered separately come to 25.50 CAD, so the page states a saving of 3.00 CAD. Plate cost for the bundle is 3.29 CAD.
Children's plates for Cedar Street Tacos. A one page children's menu: the same 4 dishes as smaller portions at 60 percent of the menu price, rounded to the nearest 0.05. The prices are 8.10, 6.60, 8.70, 7.20 CAD, and every one already lands on a five cent step, so no row on the page is rounded.
Printed menu for Willow Event Kitchen. A one page food menu with 2 sections (Buffet service, Delivered service) and 4 dishes, prices right aligned in CAD and an allergen number after every dish. First page text begins "Willow Event Kitchen". Every price is the menu_price_cad in plate-costs.csv and every allergen number is a Y column in allergen-matrix.csv.
Allergen menu for Willow Event Kitchen. A landscape page carrying 4 dishes against the 14 declarable allergens, Y for declared and a hyphen for not declared, with a second table naming the ingredient behind every Y. The grid reproduces allergen-matrix.csv in the costing folder cell for cell.
Drinks list for Willow Event Kitchen. A one page list with 4 lines by the glass and 4 by the bottle. Where a drink is sold both ways the glass price is the bottle price divided by 5 pours and rounded up to the next 0.25, which the page states. Drinks sit outside the shared ingredient costing study, so no plate cost is claimed for any of them.
Daily specials insert for Willow Event Kitchen. A dated single page for service on 2026-08-28, inside the costing period 2026-08-20 to 2026-09-02. 2 dishes from the printed menu at 85 percent of the menu price rounded to the nearest 0.05, with the menu price and the saving shown beside each.
Two portion buffet package for Willow Event Kitchen. A fixed price page at 35.00 CAD covering 2 dishes that are already on the printed menu. The same items ordered separately come to 39.00 CAD, so the page states a saving of 4.00 CAD. Plate cost for the bundle is 3.20 CAD.
Printed menu for Harbour Kitchen Group. A one page food menu with 2 sections (Burgers, Sides and salads) and 4 dishes, prices right aligned in CAD and an allergen number after every dish. First page text begins "Harbour Kitchen Group". Every price is the menu_price_cad in plate-costs.csv and every allergen number is a Y column in allergen-matrix.csv.
Allergen menu for Harbour Kitchen Group. A landscape page carrying 4 dishes against the 14 declarable allergens, Y for declared and a hyphen for not declared, with a second table naming the ingredient behind every Y. The grid reproduces allergen-matrix.csv in the costing folder cell for cell.
Drinks list for Harbour Kitchen Group. A one page list with 4 lines by the glass and 4 by the bottle. Where a drink is sold both ways the glass price is the bottle price divided by 5 pours and rounded up to the next 0.25, which the page states. Drinks sit outside the shared ingredient costing study, so no plate cost is claimed for any of them.
Daily specials insert for Harbour Kitchen Group. A dated single page for service on 2026-08-28, inside the costing period 2026-08-20 to 2026-09-02. 2 dishes from the printed menu at 85 percent of the menu price rounded to the nearest 0.05, with the menu price and the saving shown beside each.
Burger and fries set for Harbour Kitchen Group. A fixed price page at 26.50 CAD covering 2 dishes that are already on the printed menu. The same items ordered separately come to 30.00 CAD, so the page states a saving of 3.50 CAD. Plate cost for the bundle is 3.87 CAD.
Children's plates for Harbour Kitchen Group. A one page children's menu: the same 4 dishes as smaller portions at 60 percent of the menu price, rounded to the nearest 0.05. The prices are 11.10, 11.70, 6.90, 7.20 CAD, and every one already lands on a five cent step, so no row on the page is rounded.
The exact answer for the blur variant of this delivery: every one of the 29 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the clean variant of this delivery: every one of the 29 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the lowcontrast variant of this delivery: every one of the 29 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the noise variant of this delivery: every one of the 29 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the rot+3 variant of this delivery: every one of the 29 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the rot-7 variant of this delivery: every one of the 29 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the blur variant of this invoice: every one of the 33 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the clean variant of this invoice: every one of the 33 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the lowcontrast variant of this invoice: every one of the 33 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the noise variant of this invoice: every one of the 33 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the rot+3 variant of this invoice: every one of the 33 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the rot-7 variant of this invoice: every one of the 33 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the blur variant of this receipt: every one of the 23 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the clean variant of this receipt: every one of the 23 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the lowcontrast variant of this receipt: every one of the 23 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the noise variant of this receipt: every one of the 23 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the rot+3 variant of this receipt: every one of the 23 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
The exact answer for the rot-7 variant of this receipt: every one of the 23 words with its string and pixel box, plus the transform that produced this variant. Authored before the image existed, so it is ground truth by construction rather than a second opinion.
Video with a SAMPLE caption burned into the pixels (not a soft subtitle track), for OCR/caption-burn detection.
On-screen countdown digits burned into frames, for OCR and slate detection.
Hard-burned caption twin of the soft-sub clip, for burn-in detection and OCR.
On-screen running timecode burned into frames, for OCR and slate tools.
Hard-burned EN SAMPLE caption for OCR language tests.
Hard-burned ES SAMPLE caption for OCR language tests.
A single multi-turn chat example for an OCR assistant: system/user/assistant roles in JSONL.
