SynthDoc

SynthDoc is a synthetic document-page dataset with one pixel-exact evidence box per image. Each row is a question, a rendered document page that answers it, the exact sentences on that page that carry the answer, a single bounding box drawn around those sentences, and a teacher relevance score.

The boxes are not annotations. They are read straight out of the browser layout engine that drew the glyphs, so they are correct by construction — no OCR, no human labelling, no model in the loop.

Rows	5500 (5000 / 250 / 250)
Source corpora	FEVEROUS, OTT-QA, Qasper, RepLiQA
Boxes per image	exactly 1
Evidence share of page text	median 0.47
Box area	median 25.5% of the page
Constant-box baseline	IoU 0.532
Images	included in the parquet, viewable above
Page templates	6 sizes × 8 layouts × 16 themes × 13 fonts

---

Motivation

We built SynthDoc to fix a measured failure in a multimodal reranker we are distilling. The student has to emit an evidence box along with its relevance score, and on document-heavy evaluation sets its boxes came out 3.2× too tall — on docvqa the mean predicted-to-gold area ratio was 2.23. The model had learned roughly where to look but not how far the evidence extends.

The reason is in the supervision. Document VQA boxes are drawn by annotators around a region they judged relevant, at whatever granularity felt natural. That signal is noisy, inconsistent between annotators, and it never tells the model where the evidence actually stops.

SynthDoc inverts the problem. Instead of finding boxes in existing document images, it starts from text corpora where the evidence span is already known and renders them into document pages, keeping the mapping from evidence text to on-page geometry. The layout engine reports the exact rectangles it drew each evidence word into, so the box is the evidence — to the pixel.

That gives three things annotation cannot:

• Exact extent. The box ends where the evidence ends, every time.
• Controlled variation. Page size, column count, layout, theme, font, type size and table styling are sampled independently, so a model cannot reach the box through a layout shortcut.
• Text alongside geometry. Every row carries the page's plain text and the evidence sentences verbatim, so the same example supervises text grounding and visual grounding at once.

Why the evidence is a large box on a page that is still full of text

This is the design decision that matters most, and it is easy to get wrong.

A box is required to cover 20–75% of the page. The naive way to satisfy that is to put almost nothing else on the page, and it is a trap: we built that version first and measured it. Evidence was 84% of the page text, and a model that ignores the image and the query and always emits the same rectangle scored IoU 0.66. It also teaches a model that the answer is always a big box, which is the failure we set out to fix.

The fix is not a smaller box. Box area is roughly

box_area  ~  evidence_share  x  (1 - 2*margin)^2  x  page_fill

so a large box on a text-heavy page needs two things that have nothing to do with deleting content: a long evidence span (≥ 600 characters, which is why the corpora above were filtered) and tight page geometry (2–5% margins, text covering 88–98% of the usable area). With those, a 600–1,500 character evidence span reaches 20–75% of the page while the page still carries a few thousand characters of other text around it.

Geometry alone does not enforce this. An earlier build passed every other gate while evidence drifted to a median 59% of the page text, with a 90th percentile of 92% — pages that were technically valid and substantially all evidence. The renderer trims content to fit the page, which discards the dilution text added at sampling time, so the constraint has to be checked after rendering. It now is: a render is rejected if evidence exceeds 55% of the page's characters.

What the shipped data looks like:

	value
Median box area	25.5% of the page
Median evidence share of page text	0.47
Median page text	1672 characters
Constant-box mean IoU	0.532

One caveat stated plainly: IoU between two large boxes is high by construction, so the constant-box number is not comparable across datasets with different box sizes. The meaningful guarantee here is that most of every page is text the model must rule out.

---

Dataset Creation

1. Sources

Four text corpora, chosen because each already marks its evidence, each stresses a different document shape, and each has documents whose evidence span is long enough to fill a real region of a page.

corpus	what a document is	evidence unit
OTT-QA	a Wikipedia table plus its linked passages	the passage / table row grounding the answer
RepLiQA	a synthetic news / reference article about invented entities	the annotated answer sentence
FEVEROUS	a Wikipedia page verifying a claim	the sentences, list items and table cells cited
Qasper	a full NLP research paper	the sentences an annotator cited as evidence

RepLiQA matters for a specific reason: its entities do not exist, so a model cannot answer from parametric memory. It has to read the page.

Only documents whose evidence span is at least 600 characters are used. That is what makes a large box compatible with a dense page — see below.

Three corpora were tested and rejected. MuSiQue's multi-hop evidence sits in different paragraphs separated by distractors, so a single union box swallows the distractors: median box precision 0.33, 12% of renders passing. ConditionalQA has only 40 documents with a long enough evidence span. RAGBench stores its finance tables as serialized Python lists, which render as source code rather than prose. Any document containing serialized text is dropped corpus-wide by the same rule.

2. Rendering

Each document is laid out as HTML and CSS and screenshotted with headless Chromium. Before the screenshot, every evidence word is measured with the DOM Range.getClientRects() API — the same rectangles the browser used to paint the glyphs. The union of those rectangles is the box.

Style is sampled per render across 6 page geometries (A4, Letter, slide, poster, infographic, wide), 8 layouts (flow, cards, boxed, bands, sidebar, numbered, ruled, tiles), 16 colour themes, 13 open-licence typefaces, four table styles, one or two columns, and continuous line-height, margin and type-size axes. Type size is solved closed-form so the text fills a target fraction of the page rather than being picked and hoped for.

Ornament that is not content — section numbers, bullets, rules — is drawn with CSS counter() and ::before, never as DOM text, so it can never leak into the extracted page text or into a box.

3. Quality gates — code only, no human review

A render is discarded unless it passes all of:

gate	requirement
box_area	the box covers 20–75% of the page
box_precision	≥ 0.85 of the text inside the box is evidence
evidence_not_dominant	evidence is ≤ 55% of the page text (≤ 70% for FEVEROUS)
page_filled	≥ 0.55 of the usable page carries content
font_loaded	the sampled webfont actually rendered
not_clipped	no content overflowed the page
evidence_text_exact	text recovered from the boxes equals the source evidence
rows_cover_blocks	every text block is accounted for in the geometry
boxes_sane	ordered, in-bounds, non-degenerate
one_evidence_box	exactly one box survives
image_dimensions	the PNG on disk matches the recorded size
image_nonblank	the page is not blank

Roughly one render in four survives. The box_precision gate is the important one: with a single box per image, recall is 1.0 by construction, so precision is the only thing that can go wrong, and it fails exactly when the evidence is scattered enough that the union box has to span text that is not evidence. Every shipped row is above 0.85, and the per-row value is in the data.

FEVEROUS gets a looser density ceiling (70% rather than 55%), and this is a real inconsistency, not a tuned improvement. Its documents are short claim-verification pages whose cited sentences are inherently most of the text, so the global ceiling admitted under 2% of its renders. The choice was to drop the corpus or to raise its ceiling; we raised it, because table-and-list verification pages are a document shape the other three corpora do not cover. FEVEROUS rows are consequently denser than the rest — median evidence share 0.58 against 0.42–0.48 elsewhere — and there are fewer of them. generation_setting.evidence_char_fraction carries the per-row value, so this is filterable if it matters for your use.

4. Boxes are padded

The raw box is the exact glyph extent, which visually shaves the ascenders and the outer stems off the text it marks. Every shipped box is grown by 6 CSS pixels on each side and clamped to the page. generation_setting.box_tight keeps the unpadded box, and the 20–75% area gate is applied to that one.

5. Splits

Split on the source document, not on the rendered image. Several pages can be rendered from the same document, so splitting on images would leak the text across splits. Each split is balanced across OTT-QA, RepLiQA and Qasper by construction; FEVEROUS contributes about half as many rows, for the yield reason given above.

split	rows	source documents	FEVEROUS	OTT-QA	Qasper	RepLiQA
train	5000	2283	697	1435	1434	1434
val	250	165	63	63	62	62
test	250	203	63	63	62	62
Document overlap between any two splits is zero.

6. Teacher scores

Every one of the 5500 rows carries a teacher score. The score is P(yes) / (P(yes) + P(no)) from Qwen/Qwen3-VL-Reranker-8B run as a single-logit cross encoder, given the query on one side and both the page image and the page text on the other, under the reranker's own template with the instruction "Find a screenshot that relevant to the user's question." This is the identical recipe used for the teacher scores in the rest of our distillation pipeline, so SynthDoc rows can be mixed into that training mixture unchanged.

split	min	p10	median	p90	max	mean
train	0.017	0.123	0.578	0.911	0.977	0.551
val	0.025	0.125	0.746	0.912	0.962	0.594
test	0.012	0.143	0.715	0.908	0.970	0.592

Every pair here is positive by construction — the page provably contains the answer — yet the teacher spreads them across the whole [0, 1] range. That is the point: the score is a soft target, not a label. It records how confident this particular teacher is on this particular page, and the low tail is informative in its own right, marking pages where the evidence is buried in a dense layout or phrased far from the query's wording.

---

Examples

The red rectangle is bbox_2 exactly as shipped.

FEVEROUS — tiles layout on infographic

Query. Including the fact that the kingdom had no reciprocal export trade and its once-thriving industries such as shipbuilding were in deep decline; goods that were in demand had to be bought from England for sterling, the late 17th century was a difficult period for Scotland which included the 1690s, Scotland's coldest decade in the past 750 years, and its economy was relatively small, its range of exports very limited and it was in a weak position in relation to England.

Evidence inside the box.

• The late 17th century was a difficult period for Scotland, as it was for much of Europe; the years 1695-97 saw catastrophic famine in present-day Estonia, Finland, Latvia, Norway and Sweden, plus an estimated two million deaths in France and Northern Italy. The 1690s were Scotland's coldest decade in the past 750 years as documented in tree ring records. Scotland's economy was relatively small, its range of exports very limited and it was in a weak position in relation to England, its powerful neighbor (with which it was in personal union, but not yet in political union). The kingdom had no reciprocal export trade and its once-thriving industries such as shipbuilding were in deep decline; goods that were in demand had to be bought from England for sterling.

bbox_2 [0.54, 0.14, 0.9388, 0.7165] · box covers 23% of the page · precision 1.00 · teacher score 0.8467

OTT-QA — tiles layout on slide

Query. What is the population of the microstate African country with 115 islands ?

Evidence inside the box.

• Seychelles (/selz/ (listen); French: [sl] or [sel]), officially the Republic of Seychelles (French: République des Seychelles; Creole: La Repiblik Sesel), is an archipelago country in the Indian Ocean. The capital of the 115-island country, Victoria, lies 1,500 kilometres (932 mi) east of mainland Africa. Other nearby island countries and territories include Comoros, Mayotte (territory of France), Madagascar, Réunion (territory of France), and Mauritius to the south; as well as the Maldives and Chagos Archipelago to the east. With a population of roughly 94,367, it has the smallest population of any sovereign African country.

bbox_2 [0.5127, 0.1925, 0.9747, 0.7627] · box covers 26% of the page · precision 1.00 · teacher score 0.8782

Qasper — flow layout on slide

Query. What are the traditional methods to identifying important attributes?

Evidence inside the box.

• In, Pasca et al. firstly extract potential class-attribute pairs using linguistically motivated patterns from unstructured text including query logs and query sessions, and then score the attributes using the Bayes model. In, Rahul Rai proposed to identify product attributes from customer online reviews using part-of-speech(POS) tagging patterns, and to evaluate their importance with several different frequency metrics. In, Lee et al. developed a system to extract concept-attribute pairs from multiple data sources, such as Probase, general web documents, query logs and external knowledge base, and aggregate the weights from different sources into one consistent typicality score using a Ranking SVM model.
• In, Li et al. introduced the OntoRank algorithm for ranking the importance of semantic web objects at three levels of granularity: document, terms and RDF graphs. The algorithm is based on the rational surfer model, successfully used in the Swoogle semantic web search engine.

bbox_2 [0.0214, 0.589, 0.9786, 0.8264] · box covers 23% of the page · precision 0.91 · teacher score 0.3619

RepLiQA — tiles layout on letter

Query. In the account of Megan Jacobs, what unusual phenomenon did she experience at Grimsley House?

Evidence inside the box.

• Among the most enduring of such legends is that of the Grimsley House, a Victorian mansion that has stood since October 1883 It is said to be home to the spirits of its original inhabitants, the Grimsley family Witnesses have reported inexplicable footsteps, ghostly apparitions in period attire, and even the faint sounds of a piano playing from the desolate ballroom In the decades since it became an infamous landmark, generations of Fairwich residents have added to the house's mystique with stories of their own Megan Jacobs, a lifelong resident who experienced an unexplained encounter at Grimsley House in September 2023, recounted, "The air was heavy, and then, out of nowhere, a chill just ran down my spine, and I heard a whisper calling my name, though there was no one there

bbox_2 [0.5502, 0.1908, 0.9357, 0.8308] · box covers 25% of the page · precision 1.00 · teacher score 0.9171

---

Schema

Every split is a parquet table with one row per rendered page.

field	type	meaning
id	string	unique row id, also the image filename stem
source	string	which corpus the text came from
query	string	the question
answer	string	the short answer, "" when the corpus has none
text	string	the full plain text of the page, blocks joined by a blank line
image	image	the rendered page PNG, embedded — renders in the viewer
score	float	teacher relevance in [0, 1], P(yes)/(P(yes)+P(no))
bbox_2	list[list[float]]	the box. One entry, [x0, y0, x1, y1]
sentences	list[string]	the evidence text inside the box, one entry per evidence group
target	string	JSON training target, see below
evidence	string	constant "score+bbox_2d+text", names the target format
size	list[int]	[width, height] of the image in pixels
image_path	string	original path within the build directory
layout page columns theme font table_style		the sampled style, promoted for filtering
box_area	float	box area as a fraction of page area
box_precision	float	fraction of the text inside the box that is evidence
source_doc_id	string	the source document — the split key
generation_setting	string	JSON blob with the full render provenance

How the box is represented

bbox_2 holds one box, [x0, y0, x1, y1], in normalised coordinates: each value is a fraction of the image dimension, so x values divide by size[0] and y values by size[1]. Origin is the top-left corner, x grows right, y grows down. To get pixels:

x0, y0, x1, y1 = row["bbox_2"][0]
w, h = row["size"]
px = [x0 * w, y0 * h, x1 * w, y1 * h]

There is exactly one box per image, always. When the evidence is several scattered clauses, the box is their union and sentences has one entry per clause.

The target string

target is a JSON string — the literal text a generative student is trained to emit:

{"score": 0.94, "bbox_2d": [95, 258, 899, 827], "text": ["..."]}

bbox_2d is the same box as bbox_2 but rescaled to a 0–1000 integer grid (round(v * 1000)), which is the convention Qwen-VL models use for coordinates. score mirrors the score column and text mirrors sentences.

---

Usage

from datasets import load_dataset

ds = load_dataset("shredder-31/SynthDoc", split="train")
row = ds[0]
row["image"]        # PIL.Image, already decoded
row["bbox_2"][0]    # [x0, y0, x1, y1], normalised

Drawing the box:

from PIL import ImageDraw

im = row["image"].copy()
w, h = im.size
x0, y0, x1, y1 = row["bbox_2"][0]
ImageDraw.Draw(im).rectangle([x0 * w, y0 * h, x1 * w, y1 * h],
                             outline="red", width=5)

---

Composition

Layout and page geometry over the training split.

Layout

value	rows	share
bands	591	11.8%
boxed	549	11.0%
cards	575	11.5%
flow	838	16.8%
numbered	808	16.2%
ruled	801	16.0%
sidebar	641	12.8%
tiles	197	3.9%

Page geometry

value	rows	share
a4	1107	22.1%
infographic	831	16.6%
letter	1283	25.7%
poster	693	13.9%
slide	552	11.0%
wide	534	10.7%

---

Limitations

• Pages are synthetic. They are clean, digital-born renders. There is no scan noise, skew, compression artefact, handwriting or photograph. A model trained only on SynthDoc will not have seen a degraded document.
• One box only. Scattered evidence is merged into a union box, so a small amount of non-evidence text falls inside it. box_precision records exactly how much, per row, and never drops below 0.85. Corpora whose evidence is spread too far — MuSiQue, for instance — cannot be represented at all.
• The page is a window, not a whole document. Long sources are trimmed to what fits one page around the evidence, so multi-page reasoning is out of scope.
• English only, and the typefaces are Latin-script.

Licensing and provenance

Text is derived from OTT-QA (Wikipedia, CC BY-SA), RepLiQA (CC BY 4.0), FEVEROUS (CC BY-SA) and Qasper (CC BY 4.0); please observe each source's terms. Typefaces are SIL Open Font License. The renders and annotations in this repository are released under CC BY 4.0.