Datasets:

vlmgrounding
/

hardvqa-sft-dpo

source string	image_id string	category string	question string	answer string	answer_type string	adversarial_premise string	evidence_rationale string	supporting_description string	supporting_ocr string	supporting_answers string
chartqa	chartqa_237	nonexistent_object	Did state and local governments collect exactly 30 billion dollars through public utility tax in 2018?	no	boolean	The question claims a different tax collection amount than the actual.	The metadata states the amount collected was 27.48 billion, not 30 billion.	How much money did state and local governments collect through public utility tax in 2018?		27.48
chartqa	chartqa_237	chart_hallucination	Was the public utility tax collected by state and local governments in 2018 more than 50 billion dollars?	no	boolean	The question exaggerates the tax collection amount.	The metadata specifies the amount as 27.48 billion, which is less than 50 billion.	How much money did state and local governments collect through public utility tax in 2018?		27.48
chartqa	chartqa_178	nonexistent_object	Does China host the highest number of supplier factories for C&A?	no	boolean	The question incorrectly states the country with the highest number of supplier factories.	Metadata confirms Brazil hosts the highest number, not China.	What country hosts the highest number of supplier factories for C&A?		Brazil
chartqa	chartqa_178	chart_hallucination	Is the country with the highest number of supplier factories for C&A located in Europe?	no	boolean	The question misattributes the location of the country with the highest supplier factories.	Brazil is in South America, not Europe, as per metadata.	What country hosts the highest number of supplier factories for C&A?		Brazil
visual_genome	4248	object_counting	How many bushes is the woman standing between?	2	number	The question tests counting objects spatially related to the woman.	Metadata phrases include 'woman standing between two bushes' and 'woman standing between bushes'.	{'region_id': 5913307, 'image_id': 4248, 'phrase': 'A wall on the side of a building', 'x': 129, 'y': 245, 'width': 33, 'height': 42} \| {'region_id': 5913306, 'image_id': 4248, 'phrase': 'A wall on the side of a building', 'x': 129, 'y': 245, 'width': 33, 'height': 42} \| {'region_id': 5913305, 'image_id': 4248, 'phrase...
visual_genome	4248	object_attribute	Is the woman wearing a coat with large red dots on a white background?	yes	boolean	The question challenges recognition of clothing pattern and color.	Metadata states 'vinyl coat with large red dots on a white background' and 'the coat has red polka dots'.	{'region_id': 5913307, 'image_id': 4248, 'phrase': 'A wall on the side of a building', 'x': 129, 'y': 245, 'width': 33, 'height': 42} \| {'region_id': 5913306, 'image_id': 4248, 'phrase': 'A wall on the side of a building', 'x': 129, 'y': 245, 'width': 33, 'height': 42} \| {'region_id': 5913305, 'image_id': 4248, 'phrase...
docvqa	docvqa_219	ocr_hallucination	Did John Smith complete this request form?	no	boolean	The question falsely claims a different person completed the form.	OCR text clearly states 'Completed By: Lisa Drum'.	Who completed this request form?	{"page": 1, "clockwise_orientation": 359.7, "width": 1692, "height": 2245, "unit": "pixel", "lines": [{"bounding_box": [556, 89, 1002, 90, 1002, 131, 555, 131], "text": "RESPONSE CODE REQUEST", "words": [{"bounding_box": [561, 90, 736, 90, 736, 131, 562, 132], "text": "RESPONSE", "confidence": null}, {"bounding_box": [...	Lisa Drum
docvqa	docvqa_219	document_hallucination	The request form was completed by Lisa Drum.	yes	boolean	The question verifies the identity of the form completer.	OCR text explicitly shows 'Completed By: Lisa Drum'.	Who completed this request form?	{"page": 1, "clockwise_orientation": 359.7, "width": 1692, "height": 2245, "unit": "pixel", "lines": [{"bounding_box": [556, 89, 1002, 90, 1002, 131, 555, 131], "text": "RESPONSE CODE REQUEST", "words": [{"bounding_box": [561, 90, 736, 90, 736, 131, 562, 132], "text": "RESPONSE", "confidence": null}, {"bounding_box": [...	Lisa Drum
chartqa	chartqa_219	nonexistent_object	Did the total European non-life premiums amount to 500 billion in 2019?	no	boolean	The question exaggerates the total premiums amount.	Metadata states the amount as 409.25 billion, not 500 billion.	How much did the total European non-life premiums amount to in 2019?		409.25
chartqa	chartqa_219	chart_hallucination	Was the total European non-life premiums less than 300 billion in 2019?	no	boolean	The question understates the total premiums amount.	Metadata specifies 409.25 billion, which is more than 300 billion.	How much did the total European non-life premiums amount to in 2019?		409.25
visual_genome	3975	object_counting	Are there more than five blue and white chairs visible in the room?	yes	boolean	The question tests counting of specific colored chairs.	Metadata mentions many blue and white chairs grouped together and multiple blue and white chairs with coordinates.	{'region_id': 5127536, 'image_id': 3975, 'phrase': 'the roof of the room is triangle', 'x': 4, 'y': 2, 'width': 427, 'height': 277} \| {'region_id': 5127537, 'image_id': 3975, 'phrase': 'a lamp on a table ', 'x': 113, 'y': 86, 'width': 53, 'height': 66} \| {'region_id': 5127538, 'image_id': 3975, 'phrase': 'a vase on a t...
visual_genome	3975	object_attribute	Is the lamp on the table a Victorian style standing lamp?	yes	boolean	The question challenges recognition of lamp style.	Metadata includes 'victorian style standing lamp in corner' and 'victorian age lamp on table'.	{'region_id': 5127536, 'image_id': 3975, 'phrase': 'the roof of the room is triangle', 'x': 4, 'y': 2, 'width': 427, 'height': 277} \| {'region_id': 5127537, 'image_id': 3975, 'phrase': 'a lamp on a table ', 'x': 113, 'y': 86, 'width': 53, 'height': 66} \| {'region_id': 5127538, 'image_id': 3975, 'phrase': 'a vase on a t...
textcaps_or_textvqa	textvqa_325	ambiguous_evidence	Does this image contain negative number readings?	yes	boolean	The question asks about presence of negative numbers in the image.	Metadata answer indicates 'yes' multiple times, but no OCR or image text is provided to verify.	does this have negative number readings?		yes \| al gore \| yes \| yes \| yes \| yes \| yes \| yes \| yes \| yes
textcaps_or_textvqa	textvqa_325	ocr_hallucination	There are no negative number readings in this image.	no	boolean	The question contradicts the metadata answer.	Metadata answer states 'yes' for negative number readings, so the claim is false.	does this have negative number readings?		yes \| al gore \| yes \| yes \| yes \| yes \| yes \| yes \| yes \| yes
docvqa	docvqa_211	ocr_hallucination	Is the rating chosen in this document (1) = Highest?	no	boolean	The question tests correct interpretation of rating selection.	OCR text shows rating selected is (3) or 3, not (1).	What is the Rating chosen in this document?	{"page": 1, "clockwise_orientation": 359.95, "width": 1764, "height": 2257, "unit": "pixel", "lines": [{"bounding_box": [1440, 38, 1710, 41, 1709, 79, 1439, 76], "text": "SILBER GELD", "words": [{"bounding_box": [1445, 39, 1591, 41, 1591, 78, 1446, 77], "text": "SILBER", "confidence": null}, {"bounding_box": [1601, 41,...	(3) \| 3
docvqa	docvqa_211	document_hallucination	The rating chosen in this document is (3).	yes	boolean	The question verifies the rating selected.	OCR text and metadata confirm the rating is (3) or 3.	What is the Rating chosen in this document?	{"page": 1, "clockwise_orientation": 359.95, "width": 1764, "height": 2257, "unit": "pixel", "lines": [{"bounding_box": [1440, 38, 1710, 41, 1709, 79, 1439, 76], "text": "SILBER GELD", "words": [{"bounding_box": [1445, 39, 1591, 41, 1591, 78, 1446, 77], "text": "SILBER", "confidence": null}, {"bounding_box": [1601, 41,...	(3) \| 3
textcaps_or_textvqa	textvqa_239	ocr_hallucination	The poster contains the text 'Welcome to the bank'.	no	boolean	The question falsely claims specific text on the poster.	Metadata answers do not mention this text; it is not visible.	what is on the poster?		bcr \| bgr \| bcr \| unanswerable \| laissez ..tre banque gerer .s economies. \| answering does not require reading text in the image \| bcr \| perdu dons vos comptes? bcr. aissez tre banque gérer s economies. \| answering does not require reading text in the image \| banker
textcaps_or_textvqa	textvqa_239	ambiguous_evidence	The poster shows the logo 'bcr'.	yes	boolean	The question tests recognition of logo text.	Metadata answers include 'bcr' multiple times, indicating presence of this text.	what is on the poster?		bcr \| bgr \| bcr \| unanswerable \| laissez ..tre banque gerer .s economies. \| answering does not require reading text in the image \| bcr \| perdu dons vos comptes? bcr. aissez tre banque gérer s economies. \| answering does not require reading text in the image \| banker
textcaps_or_textvqa	textvqa_18	ocr_hallucination	The adventure begins on July 1, 1998.	no	boolean	The question states an incorrect adventure start date.	Metadata answers consistently state June 29, 1998, not July 1.	when does the adventure begin?		june 29, 1998 \| june 29, 1998 \| june 29, 1998 \| june 29,1998 \| june 29 1998 \| june 29, 1998 \| june 29, 1998 \| june 28, 1998 \| june 29 1998 \| june 29, 1998
textcaps_or_textvqa	textvqa_18	ambiguous_evidence	The adventure begins on June 29, 1998.	yes	boolean	The question verifies the adventure start date.	Metadata answers confirm June 29, 1998 as the start date.	when does the adventure begin?		june 29, 1998 \| june 29, 1998 \| june 29, 1998 \| june 29,1998 \| june 29 1998 \| june 29, 1998 \| june 29, 1998 \| june 28, 1998 \| june 29 1998 \| june 29, 1998

End of preview. Expand in Data Studio

HardVQA / GroundVQA

Default viewer split: 3,877 GroundVQA pairs after Tier A filtering and yes-prefix corrections.

Images are embedded in groundvqa_3877.parquet (image column, Hugging Face Image type), so the dataset viewer can show thumbnails. Private Colab / Google Drive image_path values are not stored. The following keys were also dropped: id, difficulty, review_status.

The standalone dpo_images/ folder was removed from this repo because the WearVQA SFT/DPO parquet already embeds those photos.

Hugging Face repo files

Path on this repo	What it is
`groundvqa_3877.parquet`	3,877 GroundVQA rows with embedded images (default / `groundvqa` config)
`data/train-00000-of-00001.parquet`	20 WearVQA SFT/DPO rows with embedded images (`sft_dpo` config)
`Hard_dpoVQA.jsonl`	Original WearVQA DPO jsonl
`hardVQA_dpo_example.pdf`	DPO review PDF
`README.md`	This card

Source files (local)

GroundVQA draft and filtered manifests

Role	Path	Rows
Input draft	`G:\My Drive\groundvqa1000\llm_generation\groundvqa_llm_draft_v2.jsonl`	4,238
Same draft (working copy)	`C:\Users\qinwu\groundVQA\groundvqa_llm_draft_v2.jsonl`	4,238
After Tier A filter	`C:\Users\qinwu\groundVQA\groundvqa_llm_draft_v2_filtered.jsonl`	3,877 kept
Dropped by Tier A	`C:\Users\qinwu\groundVQA\groundvqa_llm_draft_v2_dropped_tierA.jsonl`	361
Tier A count report	`C:\Users\qinwu\groundVQA\groundvqa_llm_draft_v2_tierA_report.json`	—
After yes-prefix correction (parquet source)	`C:\Users\qinwu\groundVQA\groundvqa_llm_draft_v2_filtered_corrected.jsonl`	3,877
Yes-prefix changelog	`C:\Users\qinwu\groundVQA\groundvqa_llm_draft_v2_yesprefix_changelog.jsonl`	160 rewritten answers
Removed-row copy	`C:\Users\qinwu\groundVQA\groundvqa_tierA_removed.jsonl`	361
Yes-prefix rewritten subset	`C:\Users\qinwu\groundVQA\groundvqa_yesprefix_corrected_160.jsonl`	160
Filter notes	`C:\Users\qinwu\groundVQA\filter.md`	—
Filter script	`C:\Users\qinwu\groundVQA\filter_tier_a.py`	—
Correction script	`C:\Users\qinwu\groundVQA\correct_yes_prefix.py`	—

Image roots

Original Colab field (removed from parquet):

/content/drive/MyDrive/groundvqa1000/images/<source>/<image_id>.jpg

Local image roots:

Source	Image directory
ChartQA	`G:\My Drive\groundvqa1000\images\chartqa\` and `C:\Users\qinwu\groundVQA\images\chartqa\`
DocVQA	`G:\My Drive\groundvqa1000\images\docvqa\` and `C:\Users\qinwu\groundVQA\images\docvqa\`
GQA	`G:\My Drive\groundvqa1000\images\gqa\` and `C:\Users\qinwu\groundVQA\images\gqa\`
TextCaps / TextVQA	`G:\My Drive\groundvqa1000\images\textvqa\` and `C:\Users\qinwu\groundVQA\images\textvqa\`
Visual Genome	`G:\My Drive\groundvqa1000\images\visual_genome\` and `C:\Users\qinwu\groundVQA\images\visual_genome\`
WearVQA DPO	`C:\Users\qinwu\groundVQA\images\dpo_images_wearvqa_20\` (uploaded as `dpo_images/`)

Parquet image_file format after stripping the private prefix:

images/<source>/<filename>

Examples:

images/chartqa/chartqa_237.jpg
images/docvqa/docvqa_13.jpg
images/gqa/gqa_240.jpg
images/textvqa/textvqa_363.jpg
images/visual_genome/2733.jpg

Filter procedures

Applied 2026-09-13. Only Tier A plus a yes-prefix rewrite were applied. Tier B quarantine and Tier C human audit were planned, not run as automatic drops.

Step 1 — Input

groundvqa_llm_draft_v2.jsonl
4,238 rows

Source mix before filter: Visual Genome 1,966; TextCaps/TextVQA 810; DocVQA 590; ChartQA 548; GQA 324.

Step 2 — Tier A automatic drop (`filter_tier_a.py`)

Drop a row if any rule matches.

ID	Rule	Hits (a row can match more than one)
A1	Answer matches `cannot` / `can not` / `can't` determine (or close variants)	311
A2	Ungrounded meta-question: visible objects/text, “to answer the question”, “clear evidence to answer…”	34
A3	Answer is `not present` or `not visible` and the question is generic visibility/existence with no named object	48

Result: 4,238 → 3,877 kept, 361 dropped.

Kept by source: Visual Genome 1,873; TextCaps/TextVQA 730; DocVQA 585; ChartQA 541; GQA 148.

input:   groundvqa_llm_draft_v2.jsonl
output:  groundvqa_llm_draft_v2_filtered.jsonl
dropped: groundvqa_llm_draft_v2_dropped_tierA.jsonl
report:  groundvqa_llm_draft_v2_tierA_report.json

Step 3 — Yes-prefix correction (`correct_yes_prefix.py`)

Does not drop rows. Rewrites confirmation-style answers that only echoed a fact from the question:

<fact>  ->  yes, <fact>

Skipped already yes/no/true/false answers, not present / not visible, and pure How many ...? counting questions.

Result: 3,877 rows still; 160 answers rewritten.

input:      groundvqa_llm_draft_v2_filtered.jsonl
output:     groundvqa_llm_draft_v2_filtered_corrected.jsonl
changelog:  groundvqa_llm_draft_v2_yesprefix_changelog.jsonl

Step 4 — Parquet export (this repo)

input:   C:\Users\qinwu\groundVQA\groundvqa_llm_draft_v2_filtered_corrected.jsonl
output:  groundvqa_3877.parquet

Removed id, image_path, difficulty, and review_status
Embedded each photo in the image column (no Drive paths; filename only, e.g. chartqa_237.jpg)
Kept VQA fields: source, image_id, category, question, answer, answer_type, adversarial_premise, evidence_rationale, supporting fields, review notes, and yes-prefix answer_original / answer_correction