A vision-language model read each image together with its QA and judged the item. The pass is not a filter that only removes rows — it rewrites answers it finds wrong but salvageable, drops what it cannot salvage, and adds QA where the image supports more than the source provided. Each row carries the judge's own record in clean_meta, including the cleaner identity, the policy it applied, and its per-item scores for legibility, richness and coverage.

A large share of the QA here was written by the cleaning pass, not by the original dataset. Across the pool that share runs from roughly half to over 80% of a family's turns, and it is reported in the table above. Those rows have an empty org_answer because no pre-clean original exists. Treat them as model-generated supervision: they were judged against the image, but they are not human annotation, and model-written QA is where formatting defects are most likely.

The effect on the answers that were carried over is substantive rather than cosmetic. In this pool the pass corrects values, not just wording — a curve's critical point restated from 4.00 to 2.00, a computed ratio from 1 to Approximately 1.33 — and for some families it removed the content entirely, which is why those families are absent here rather than published empty.

`answer` vs `org_answer`

answer — the cleaned answer. Train on this.
org_answer — the pre-cleaning answer from the same (image, question) in the uncleaned pool. It is empty for QA the cleaning pass added, which have no original.

For this family org_answer is empty throughout: recovering it means scanning the whole pre-clean family, which is raw family is 18 GB (> 6 GB cap). The cleaned answer is unaffected.

Where the two differ, the difference is the correction. Keeping both makes every edit auditable instead of asking you to take the cleaning on trust.

Schema

image — HF Image(), renders directly in the dataset viewer
image_sha256 — content hash of the image
n_qa — number of turns attached to this image
qa — list of {question, answer, org_answer}, in source order
family / clean_meta — provenance and the judge's per-row record

A row with more than one entry in qa is a conversation over a single image, not a set of independent questions — the turns depend on each other and the image belongs to all of them. Keep them together and in order when training.

Parquet row groups are written at ~32 MB so the dataset viewer pages without stalling.

What is NOT claimed

No rejection sampling was run against this family: there is no accept/reject verdict per row, and answer is the cleaned reference rather than a model generation that a string verifier confirmed. For that, use the *_RS_think / *_rejected datasets in elliot-mllm.

Examples

10 rows taken straight from this dataset. Images are the original files as stored here - no downscaling, no recompression - and every turn is shown in full.

Example 1 — 10 turns on one image

#	question	answer
1	Who hates the algorithm? Offer a terse response.	The unabomber.
2	What are these advertisements for? Your answer should be compact.	Ask and Google.
3	What is the size of the white advertisement board? Your answer should be compact.	12 x 25.
4	Relay a brief, clear account of the picture shown.	A street scene featuring two billboards: a top one for Google (800-GOOG-411) and a bottom green one for Ask stating 'The Unabomber Hates The Algorithm'.
5	Share a concise interpretation of the image provided.	A building outside has a billboard on top of it, advertising Google services.
6	Present a compact description of the photo's key features.	A green sign says that the unabomber hates the algorithm.
7	Provide a one-sentence caption for the provided image.	A street scene featuring two billboards: a top one for Google (800-GOOG-411) and a bottom green one for Ask stating 'The Unabomber Hates The Algorithm'.
8	What is the full phone number displayed on the top billboard?	800-GOOG-411
9	What identification numbers are printed on the metal frame of the top billboard?	1273 and #356713
10	What structure is attached to the side of the building to the left of the billboards?	Scaffolding

Example 2 — 3 turns on one image

#	question	answer
1	What is the title of the yellow book? Provide a short and direct response.	Fail OF THE ร้อย YEARS
2	What number is written on the white price tag attached to the top left of the yellow book?	50
3	What text is visible on the spines of the white books stacked behind the yellow book?	I ROAM ALONE

Example 3 — 2 turns on one image

#	question	answer
1	who makes this laptop?	hp
2	What model name is displayed on the screen?	Pavilion 15

Example 4 — 7 turns on one image

#	question	answer
1	what is the value for the stamp on the top right?	40
2	What is the date of issue printed on the bottom left of the envelope?	14 - 8 - 1972
3	What is the Vietnamese title printed at the top left of the envelope?	LÍNH THỦ ĐỘI XƯA
4	What is the French text below the main title?	LE SOLDAT D'AUTREFOIS
5	What is the value of the stamp in the top middle position?	10d
6	What is the value of the stamp on the bottom right?	30d
7	Which city is named in the circular postmarks?	SAIGON

Example 5 — 5 turns on one image

#	question	answer
1	what time is it?	10:56
2	What brand and model of watch is worn on the wrist?	Rolex Submariner
3	What date is displayed in the watch's date window?	21
4	Describe the bracelets worn on the wrist next to the watch.	A bracelet with light brown wooden beads and an orange and black paracord bracelet.
5	What objects are visible in the background behind the wrist?	Two open folding knives and a prickly pear cactus with yellow flowers.

Example 6 — 6 turns on one image

#	question	answer
1	what does license plate number read?	my3kyds
2	What car manufacturer logo is centered above the license plate?	BMW
3	Which state name is written in red script on the license plate?	California
4	What month and year are indicated by the stickers on the top corners of the plate?	DEC 2006
5	What text is embossed on the top of the black license plate frame?	IRVINE
6	What text is embossed on the bottom of the black license plate frame?	SADDLEBACK

Example 7 — 7 turns on one image

#	question	answer
1	Question: how many days left according to the watch? Answer with a single word or phrase.	5
2	How many days until the event?	There are five days left until the event.
3	How many days does this watch indicate are left?	The watch indicates that there are five days left.
4	How does the soft shadowing around the watch suggest the nature of the light source?	The soft shadowing indicates that the light source is diffused or not directly in line with the object, resulting in even lighting without harsh shadows. This suggests an indirect light source such as natural daylight coming through a window or artificial ambient lighting within an indoor space. The lack of distinct edges to the shadow implies the presence of multiple light sources contributing to overall illumination rather than a single direct one.
5	What icon is visible above the number 5 on the watch face?	A red heart.
6	What abbreviations for days of the week are visible on the outer ring of the watch face?	S, M, T, W, Th, F, S.
7	What color is the watch strap?	Black.

Example 8 — 9 turns on one image

#	question	answer
1	Question: what number can be seen on the box? Answer with a single word or phrase.	100
2	Question: what % satisfaction from the pizza company? Answer with a single word or phrase.	100
3	What aspect of the image quality should we focus on when assessing its clarity?	We should focus on sharpness, resolution, contrast, lighting conditions, color balance, noise level, framing, and composition aspects while assessing the overall clarity of the image.
4	How does the arrangement of pizza boxes and beer bottles suggest that people may have been socializing at this location?	The presence of multiple pizza boxes stacked together with several different types of beer bottles suggests an informal gathering or party where food and drinks were shared among attendees, indicating social interaction and leisure activities such as eating and drinking in a relaxed setting like a bar or casual dining environment.
5	What factors contribute to the popularity of takeout food options like pizza combined with beverages?	Convenience, variety, social appeal, affordability, ease of delivery or pick-up, and customization choices all contribute to the popularity of takeout food options such as pizza combined with beverages. These aspects cater to busy lifestyles, diverse tastes, and social gatherings while providing accessible dining solutions that can be enjoyed in various settings.
6	What text is printed on the side of the pizza boxes?	The text 'Producto de Consumo Inmediato' and 'Calidad 100% Garantizada' is printed on the side of the boxes.
7	What are the brands of the three beer bottles visible on the table?	The brands are Pacena (left), Beck's (middle), and Austral (right).
8	What color is the table the items are placed on?	The table is red.
9	How many pizza boxes are in the stack?	There are eight pizza boxes in the stack.

Example 9 — 7 turns on one image

#	question	answer
1	Question: what year is listed? Answer with a single word or phrase.	2010
2	What color is the label design?	The label design is primarily dark blue with white text and stars. The constellation-like pattern adds a subtle hint of lighter shades within the star designs. This combination gives it an elegant and sophisticated look that matches well with its branding as "Celeste."
3	What brand name can be seen in the label?	The brand name written on the label is TORRES.
4	What wine region is printed on the white band of the label?	Ribera del Duero
5	What text appears directly above the brand name 'TORRES'?	Selección de
6	What is the full Spanish text block located below 'CRIANZA'?	EN LAS NOCHES DE VENDIMIA, ESTAS SON LAS ESTRELLAS QUE VIGILAN Y CONTEMPLAN EL NACIMIENTO DE CELESTE
7	What small text is visible near the constellation stars on the left side of the label?	LA MINOR

Example 10 — 8 turns on one image

#	question	answer
1	Question: what page number is this page? Answer with a single word or phrase.	273
2	Question: what is the title? Answer with a single word or phrase.	unanswerable
3	What year does the handwritten note at the bottom of the page mention?	The handwritten note mentions the year "1783".
4	Provide a one-sentence caption for the provided image. Reference OCR token: XIX.AVGVS, NV., XIIII, CAL, SEPTEM., ANOE, AVGESTYS, CESAR, moritur,, cum, compleffet, 4nnum, 76.fete,, Alno, Chrifti, natalem, 16., RIDERICVS, Imperator,, Maximilianipater, anquillifSune, ritur, Chrifti, 493., atatis, veroanno, 78., duebus, maior, Augufto, Ce:, fare,, ab, eodem, Imperio, fuperatus,, 53., &, dies, totidem, b.annos, fepultus, Stephant., Cafimiri, III, Polonia, Vilna, cum, nunciATes, tur, praclara, Tartaris, victoria, per, legatum, fuum, parta, 506.7egni, regmi, An:, quinte,atatis, 45+, Promohes, sum, Arehisymnatio, Vaamah, m, Bauslaunam, formalum, SS., Thodagia, aRP, Mapim, Prearsfton, mu, oboms, ht, hmpe, 3i, famliah's, Jilubpira, Oriane, Rulic:	A page from an old printed book featuring Latin historical entries and a handwritten academic note from 1783 at the bottom.
5	What event regarding Augustus Caesar is described in the first paragraph of the printed text?	The text states that Augustus Caesar died ('moritur') in the year 16 after the birth of Christ, having completed his 76th year of age.
6	According to the text, when and where did Emperor Frederick III die?	He died in the year 1493 and was buried in Vienna in the church of St. Stephen ('Viennæ ſepultus in æde S. Stephani').
7	What academic achievement is recorded in the handwritten note at the bottom?	The note records that the writer was promoted to the Bachelor's degree in Holy Theology ('Baudalauram formatam S.S. Theologia') at the University of Vienna ('Archigymnasio Vindobonensi') in 1783.
8	What date is indicated by the Roman numerals 'XIX. AVGVS.' at the top left?	It indicates the 19th of August.

Downloads last month: 24

Total file size:

17.2 GB

You need to agree to share your contact information to access this dataset

textvqa_train_cleaned

How this was cleaned

answer vs org_answer