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.

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 — 2 turns on one image

#	question	answer
1	What is the style of the text in the image?	The text is handwritten in a cursive script.
2	How many lines of text are visible?	There are two lines of text.

Example 2 — 3 turns on one image

#	question	answer
1	Describe the text written in this photo.	In doing so he described his distinguished patient and his symptoms.
2	What is the medium and color of the text?	It is handwritten in black ink.
3	How is the text arranged spatially?	It is written across two lines.

Example 3 — 3 turns on one image

#	question	answer (cleaned)	org_answer (pre-judge)
1	Detail the handwritten content in this image.	They think there is something inevitably "queer" about psychology and this feeling of queerness usually boils down to a quite mistaken belief that the psychologist first looks into ...	They think there is something inevitably" queer" about psychology and this feeling of queerness usually boils down to a quite mistaken belief that the psychologist first looks into ...
2	What is the visual style and medium of the text in the image?	It is handwritten cursive script in black ink on a white background.	(unchanged)
3	How many lines of text are visible in the image?	There are six lines of text.	(unchanged)

Example 4 — 2 turns on one image

#	question	answer
1	What words are inscribed in this image?	The migrating stages all come from 100-250 meters, and the time of their arrival at the surface is directly related to their swimming capacity: 3rd ( oldest ) calyptopis from 1800-2200 hr; 2nd, from 2200-0200 hr; and 1st, from 0200-0600 hr.
2	What is the visual style of the text in the image?	The text is handwritten in black ink on a white background.

Example 5 — 3 turns on one image

#	question	answer
1	Output the text in this image.	I forgot to tell you, we don't usually lend any house built earlier than 1918. "
2	Is the text in the image typed or handwritten?	The text is handwritten.
3	How many lines of text are visible in the image?	There are two lines of text.

Example 6 — 3 turns on one image

#	question	answer
1	What text does this image contain?	He was here, near the railway station, three or four years.
2	What is the writing style of the text in the image?	Handwritten cursive.
3	How is the text arranged spatially?	The text is split across two lines.

Example 7 — 3 turns on one image

#	question	answer
1	Detail the handwritten content in this image.	A huge crowd had collected.
2	How is the text spatially arranged in the image?	The text is split into two parts: 'A huge' is written in the top right corner, and 'crowd had collected.' is written in the bottom left area.
3	What is the style of the writing shown?	The text is handwritten in a cursive script using black ink on a white background.

Example 8 — 3 turns on one image

#	question	answer
1	What is scribbled in this image?	Philip allowed # it to be seen that he thought going to the pictures just for the fun of it was
2	How many lines of text are visible in the image?	Three.
3	What is visible at the very beginning of the second line?	A scribbled-out or crossed-out word.

Example 9 — 4 turns on one image

#	question	answer
1	What words are inscribed in this image?	Its terms have set the Prime Minister an exacting problem.
2	What style of writing is used for the text in the image?	Handwritten cursive script.
3	What color is the ink used for the text?	Black.
4	How many lines of text are visible in the image?	Two.

Example 10 — 4 turns on one image

#	question	answer (cleaned)	org_answer (pre-judge)
1	Decode the message shown.	This emphasis on the legality of the former- Government suggests that all is not well with the political and military leadership of the pro-communists,	This emphasis on the legality of the former Government suggests that all is not well with the political and military leadership of the pro-communists.
2	How many lines of text are visible in the image?	Four	(unchanged)
3	What punctuation mark appears at the end of the first line?	A hyphen	(unchanged)
4	Is the text in the image typed or handwritten?	Handwritten	(unchanged)

Downloads last month: 26

Total file size:

1.3 GB

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

iam_train_cleaned

How this was cleaned

answer vs org_answer