Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code:   DatasetGenerationError
Exception:    CastError
Message:      Couldn't cast
sample_id: int64
dataset: string
image: string
question: string
gold_answer: string
feat_hw: list<item: int64>
  child 0, item: int64
branch: string
K: int64
K_topR: int64
n_actions: int64
reward_mean: double
reward_std: double
ev_maps_path: string
thought: string
bboxs: list<item: list<item: double>>
  child 0, item: list<item: double>
      child 0, item: double
answer: string
split: string
possible_answers: list<item: string>
  child 0, item: string
reasoning: list<item: struct<operation: string, dependencies: list<item: int64>, argument: string>>
  child 0, item: struct<operation: string, dependencies: list<item: int64>, argument: string>
      child 0, operation: string
      child 1, dependencies: list<item: int64>
          child 0, item: int64
      child 2, argument: string
width: int64
height: int64
full_answer: string
to
{'question': Value('string'), 'answer': Value('string'), 'full_answer': Value('string'), 'image': Value('string'), 'width': Value('int64'), 'height': Value('int64'), 'bboxs': List(List(Value('float64'))), 'dataset': Value('string'), 'split': Value('string'), 'reasoning': List({'operation': Value('string'), 'dependencies': List(Value('int64')), 'argument': Value('string')}), 'thought': Value('string'), 'possible_answers': List(Value('string'))}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              sample_id: int64
              dataset: string
              image: string
              question: string
              gold_answer: string
              feat_hw: list<item: int64>
                child 0, item: int64
              branch: string
              K: int64
              K_topR: int64
              n_actions: int64
              reward_mean: double
              reward_std: double
              ev_maps_path: string
              thought: string
              bboxs: list<item: list<item: double>>
                child 0, item: list<item: double>
                    child 0, item: double
              answer: string
              split: string
              possible_answers: list<item: string>
                child 0, item: string
              reasoning: list<item: struct<operation: string, dependencies: list<item: int64>, argument: string>>
                child 0, item: struct<operation: string, dependencies: list<item: int64>, argument: string>
                    child 0, operation: string
                    child 1, dependencies: list<item: int64>
                        child 0, item: int64
                    child 2, argument: string
              width: int64
              height: int64
              full_answer: string
              to
              {'question': Value('string'), 'answer': Value('string'), 'full_answer': Value('string'), 'image': Value('string'), 'width': Value('int64'), 'height': Value('int64'), 'bboxs': List(List(Value('float64'))), 'dataset': Value('string'), 'split': Value('string'), 'reasoning': List({'operation': Value('string'), 'dependencies': List(Value('int64')), 'argument': Value('string')}), 'thought': Value('string'), 'possible_answers': List(Value('string'))}
              because column names don't match
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1880, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

question
string
answer
string
full_answer
string
image
string
width
int64
height
int64
bboxs
list
dataset
string
split
string
reasoning
list
thought
string
possible_answers
list
What is on the wall?
mirror
The mirror is on the wall.
2337160.jpg
374
500
[ [ 188, 1, 252, 47 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "wall (3709487)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,on,s (2583549)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the wall within the image to focus on objects attached to it or located on it. 2. Determine the relationship of objects with the wall to find out what specific items are placed on it. 3. Once the relationship is established, seek out the names or descriptions of the items that are on the wall.
null
In which year was the cost per immigrant per day equal to $128?
2005
null
35399.jpeg
620
804
[ [ 35.63546121120453, 222.72140228748322, 268.2999658584595, 241.4519467279315 ], [ 79.1610386967659, 422.89251708984375, 145.68883001804352, 441.78631277382374 ], [ 15.686638541519642, 682.4482855796814, 564.719004817307, 696.3256161957979 ] ]
infographicsvqa
train
null
null
[ "2005" ]
what is the name of this business?
udon west
null
c86c779bfa30a5b6.jpg
1,024
768
[ [ 441.88885498046875, 18.984825611114502, 575.1322937011719, 68.26592874526978 ], [ 594.4913940429688, 24.75300693511963, 721.1162414550781, 70.77929878234863 ], [ 718.296142578125, 209.4532470703125, 805.858642578125, 233.56062984466553 ] ]
textvqa
train
null
null
null
What is on the road that is presented in this photo?
truck
The truck is on the road.
1593222.jpg
1,024
685
[ [ 128, 234, 878, 574 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "road (3361971)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,on,s (2534331)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the road in the image, as roads are common areas for vehicles and objects to be present on. 2. Determine what object or objects are on the road by looking for any items that are situated on its surface. 3. Name the object or objects found on the road to answer the question.
null
what is the website link on this van?
www.firstaid.org.uk
null
045e6da6fd7f6fb0.jpg
1,024
576
[ [ 594.831787109375, 346.68515396118164, 713.9737014770508, 361.7867159843445 ] ]
textvqa
train
null
null
null
Who is wearing a shirt?
woman
The woman is wearing a shirt.
2334133.jpg
500
375
[ [ 63, 69, 218, 375 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "shirt (3342569)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "person,wearing,s (3342568)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the item labeled as "shirt" in the image. 2. Determine which person in the image is associated with wearing the identified shirt. 3. Obtain the name or identity of the person wearing the shirt.
null
What's the chef making?
pizza
The chef is making the pizza.
2393316.jpg
500
332
[ [ 65, 122, 311, 176 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "chef (1220970)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,making,o (1220969)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the person mentioned in the question, who is referred to as 'chef'. 2. Determine the action that the chef is performing by looking for a verb that indicates an activity, such as 'making'. 3. Once the action is identified, find the object associated with this action to understand what is being made by the ch...
null
What is the % of Children in Urban Area who play at Street?
7.4
null
kzcj0227_9.png
2,556
1,826
[ [ 1409, 623, 1475, 662 ] ]
docvqa
train
null
null
[ "7.4" ]
Who is standing?
people
The people are standing.
2412949.jpg
500
362
[ [ 222, 107, 499, 173 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "person (183488)" }, { "operation": "filter pose", "dependencies": [ 0 ], "argument": "standing" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the figures labeled as 'person' within the image. 2. Determine the pose of these figures to ascertain which ones are in a standing position. 3. For the figures that are confirmed to be standing, obtain their names or identifiers to answer the question regarding who is standing.
null
What is the Total?
$15,000.
null
snkl0226_2.png
846
1,102
[ [ 559, 494, 626, 510 ], [ 559, 494, 626, 510 ] ]
docvqa
train
null
null
[ "$15,000.", "15,000." ]
What is the officer to the right of the man wearing?
helmet
The officer is wearing a helmet.
2355752.jpg
500
333
[ [ 434, 131, 442, 137 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "man (826786)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "officer,to the right of,s (826779)" }, { "operation": "relate", "dependencies": [ 1 ], "argument": "_,wearing,o (826...
1. Identify the man in the image. 2. Locate the officer who is positioned to the right of the identified man. 3. Observe what the located officer is wearing. 4. Report the item of clothing or gear that the officer is wearing on their head.
null
In which place did the 2005 shooting incident happen?
Red Lake
null
38957.jpeg
768
539
[ [ 418.94801330566406, 104.15061025321484, 481.47856521606445, 117.17500553280115 ] ]
infographicsvqa
train
null
null
[ "Red Lake" ]
what is the title mentioned in the bold letters
TENNESSEE CONSERVATION LEAGUE
null
jnfb0228_6.png
1,312
748
[ [ 337, 75, 947, 108 ], [ 69, 178, 1228, 206 ], [ 68, 368, 556, 394 ], [ 337, 75, 947, 108 ], [ 69, 178, 1228, 206 ], [ 68, 368, 556, 394 ] ]
docvqa
train
null
null
[ "Tennessee Conservation League", "TENNESSEE CONSERVATION LEAGUE" ]
What is the bunny on?
desk
The bunny is on the desk.
2357136.jpg
500
375
[ [ 1, 31, 498, 373 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "bunny (814631)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,on,o (814633)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the object referred to as "bunny" in the image. 2. Determine the relationship that indicates what the bunny is on, expressed as "bunny, on, object". 3. Use the established relationship to query the name of the object that the bunny is on.
null
Who is serving the ball?
woman
The woman is serving the ball.
2372733.jpg
330
500
[ [ 180, 183, 274, 463 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "ball (2229785)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "person,serving,s (3506037)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the object mentioned in the question, which is the ball. 2. Determine the relationship linking the ball to the action of serving, and also identify the person involved in this action. 3. Ask for the name or identifier of the person who is performing the action (serving the ball).
null
what are the letters on their shirts?
olve
null
488c7b5f4187a9ee.jpg
681
1,024
[ [ 298.9966550767422, 379.9344482421875, 422.60478338599205, 436.6667938232422 ] ]
textvqa
train
null
null
null
What percent of women own a house in the province with the highest rate of physical violence?
6.1
null
40440.jpeg
960
1,398
[ [ 505.38305282592773, 617.5440292954445, 541.1297965049744, 638.9486132115126 ], [ 422.6124572753906, 1167.8297671079636, 440.8630621433258, 1180.756151759997 ] ]
infographicsvqa
train
null
null
[ "6.1" ]
what restaurant are the protesters in front of?
ruby tuesday
null
50de37badd67a06d.jpg
1,024
768
[ [ 222.50115966796875, 40.143802642822266, 292.5679168701172, 93.38786888122559 ], [ 276.852294921875, 210.53086853027344, 329.9151420593262, 244.6005220413208 ] ]
textvqa
train
null
null
null
what is the license plate of the taxi?
d7f-621
null
1f14388c4e9189d2.jpg
1,024
765
[ [ 270.5987854003906, 401.3801684975624, 385.9912872314453, 433.3724691532552 ] ]
textvqa
train
null
null
null
What kind of furniture is in the classroom?
desk
The piece of furniture is a desk.
713627.jpg
685
1,024
[ [ 249, 384, 437, 412 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "classroom (3701796)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "furniture,in,s (1586813)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify that the setting is a classroom. 2. Locate the item of furniture within the classroom. 3. Determine the name of the item of furniture present in the classroom.
null
Who is lying in the bed?
man
The man is lying in the bed.
2342972.jpg
500
375
[ [ 0, 0, 461, 375 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "bed (3918116)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "person,lying in,s (3918124)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the bed within the scene. 2. Look for a person who is in a position that can be described as "lying in" in relation to the bed. 3. Once the person lying in the bed is identified, determine the name or the gender of the individual to describe who is lying in the bed.
null
To which company does this letterhead belong to?
William L. Strauss & Son Travel Agency
null
qkdv0228_6.png
1,775
2,267
[ [ 445, 425, 1294, 566 ], [ 445, 425, 1294, 566 ] ]
docvqa
train
null
null
[ "william l. strauss & son travel agency", "William L. Strauss & Son Travel Agency" ]
What are the "Number of Men Randomized" ?
652
null
rxvh0227_19.png
1,690
2,180
[ [ 720, 405, 785, 437 ] ]
docvqa
train
null
null
[ "652" ]
Who is the Executive Director?
Samir Kumar Modi
null
jybx0223_67.png
1,653
2,275
[ [ 820, 1968, 1051, 1992 ] ]
docvqa
train
null
null
[ "Samir Kumar Modi" ]
By whom was this document prepared?
C. Beck
null
rnjf0226_1.png
1,700
2,236
[ [ 343, 1611, 850, 1656 ] ]
docvqa
train
null
null
[ "C. Beck, LLC", "C. Beck" ]
What appliance is to the left of the woman?
stove
The appliance is a stove.
2370925.jpg
375
500
[ [ 72, 202, 123, 216 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "woman (600873)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "appliance,to the left of,s (600887)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the woman in the image. 2. Determine the relative position 'to the left of' from the perspective of the woman. 3. Locate the appliance that is situated to the left of the woman. 4. Name the appliance found to the left of the woman.
null
Who is it addressed to?
Sharon Dawson, RJR
null
ffjw0023_1.png
1,692
2,245
[ [ 566, 182, 953, 221 ] ]
docvqa
train
null
null
[ "Sharon Dawson, RJR" ]
What is located on top of the bed?
card
The card is on top of the bed.
2003.jpg
800
600
[ [ 481, 392, 552, 429 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "bed (1645781)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,on top of,s (1645775)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the location referred to as "bed". 2. Determine what is directly on top of the identified location, "bed". 3. Inquire about the name or description of the object found in the step above to ascertain what it is.
null
where is the runner about to cross?
finish
null
d0d6122f0610a663.jpg
1,024
680
[ [ 460.167724609375, 155.29776811599731, 711.1731872558594, 214.5420154929161 ] ]
textvqa
train
null
null
null
what is the ratio of State rate to U.S (100) in New York ?
73
null
lmyc0227_1.png
1,810
3,127
[ [ 1545, 989, 1592, 1023 ], [ 1541, 2120, 1600, 2154 ] ]
docvqa
train
null
null
[ "73" ]
what is the phone number mentioned in the form ?
1-800-665-6326
null
mrpx0078_6.png
1,692
2,245
[ [ 530, 970, 1206, 1008 ] ]
docvqa
train
null
null
[ "1-800-665-6326" ]
What is the projected growth rate of spending on games in Australia?
8%
null
30562.jpeg
850
1,619
[ [ 100.57617612183094, 557.9743098020554, 180.41797950863838, 593.0206259898841 ], [ 449.2701590061188, 1032.438553929329, 512.409920617938, 1073.513027112931 ] ]
infographicsvqa
train
null
null
[ "8%" ]
During which period, Lynn B. Bailey worked as Research Assistant at Clemson University?
6/70-8/72
null
zlkm0227_7.png
1,782
2,286
[ [ 1331, 1841, 1488, 1877 ] ]
docvqa
train
null
null
[ "6/70-8/72" ]
What vehicle is to the right of the bench?
car
The vehicle is a car.
2351639.jpg
500
375
[ [ 145, 150, 499, 294 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "bench (2688970)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "vehicle,to the right of,s (1790518)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the location of the bench in the image. 2. Look for the vehicle that is situated to the right of the bench. 3. Determine the type of the vehicle that is to the right of the bench.
null
What is the date given?
June 5, 1970
null
jpcd0227_1.png
1,790
2,287
[ [ 523, 880, 740, 917 ], [ 523, 880, 740, 917 ] ]
docvqa
train
null
null
[ "june 5, 1970", "June 5, 1970" ]
What cooking utensil is shiny?
tea pot
The cooking utensil is a tea pot.
2405681.jpg
375
500
[ [ 171, 42, 340, 249 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "cooking utensil (330097)" }, { "operation": "filter", "dependencies": [ 0 ], "argument": "shiny" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify all items in the image that are classified as cooking utensils. 2. From the cooking utensils identified, determine which utensil appears shiny based on its appearance in the image. 3. Provide the name of the cooking utensil that has been determined to be shiny.
null
What is in the planter?
plants
The plants are in the planter.
2405799.jpg
500
375
[ [ 0, 2, 42, 24 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "planter (4151270)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,in,s (4151269)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the object referred to as the "planter" in the image. 2. Determine what the relationship "in" signifies in this context. 3. Using the relationship "in," establish what objects or items are contained within the planter. 4. Extract the name or description of the items located inside the planter to answer the ...
null
What is the consumer's name?
Freeman Ellenberg
null
jgpw0065_2.png
1,692
2,245
[ [ 84, 431, 522, 462 ] ]
docvqa
train
null
null
[ "Freeman Ellenberg" ]
what is the auth. no. of Tyrone W Austin ?
1380
null
ymlg0227_4.png
2,648
2,284
[ [ 1593, 346, 1995, 380 ], [ 1586, 379, 1997, 413 ] ]
docvqa
train
null
null
[ "1380" ]
What's on the roof?
dome
The dome is on the roof.
2318401.jpg
251
500
[ [ 194, 292, 241, 323 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "roof (4483389)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,on,s (4483388)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify and focus on the element in the image that is described as the "roof." 2. Determine what object or structure is situated on top of the roof by establishing a relationship between the roof and any items located on it. 3. Once the relationship has been established, identify the name or description of the obje...
null
What percent of the Ireland's bathing waters are of good water quality standard?
84.4%
null
31935.jpeg
1,500
2,122
[ [ 818.3228373527527, 570.8829348683357, 902.8864558786154, 602.9216634724289 ] ]
infographicsvqa
train
null
null
[ "84.4%" ]
The program is about which subject?
Proteins In Food and Nutrition
null
rtxm0227_1.png
1,030
1,804
[ [ 177, 325, 859, 355 ], [ 323, 995, 717, 1024 ], [ 175, 1371, 864, 1409 ], [ 177, 325, 859, 355 ], [ 323, 995, 717, 1024 ], [ 175, 1371, 864, 1409 ] ]
docvqa
train
null
null
[ "proteins in food and nutrition", "Proteins In Food and Nutrition" ]
who sent a 3D bitcoin model to outer space
genesis
null
30176.jpeg
1,200
1,697
[ [ 500.4095792770386, 1547.19375872612, 637.2917175292969, 1576.8721796069294 ] ]
infographicsvqa
train
null
null
[ "genesis" ]
What is the heading of this document?
Wisdom Hall of Fame
null
zybd0227_6.png
1,780
2,314
[ [ 126, 138, 1640, 245 ], [ 538, 839, 1236, 863 ], [ 522, 929, 1605, 956 ], [ 521, 1012, 1605, 1044 ], [ 516, 1264, 1601, 1297 ] ]
docvqa
train
null
null
[ "Wisdom Hall of Fame" ]
Which of these keywords were talked about more in Japan - outbreak, lockdown or infections?
infections
null
10127.jpeg
2,550
8,400
[ [ 941.1093324422836, 1711.018717288971, 2189.188840985298, 1774.4129575788975 ], [ 812.7775266766548, 6132.034206390381, 2210.950045287609, 6199.101686105132 ], [ 330.49163594841957, 7266.098213195801, 1780.8414347469807, 7352.254534140229 ] ]
infographicsvqa
train
null
null
[ "infections" ]
Which pet food is considered as a good source of both soluble & insoluble fibre?
Beet Pulp
null
30051.jpeg
519
906
[ [ 157.63255020976067, 192.45732763409615, 216.05919279903173, 206.82106270641088 ] ]
infographicsvqa
train
null
null
[ "Beet Pulp" ]
what is the blue player that is standings number?
14
null
5824ac04cd920f0c.jpg
1,024
683
[ [ 764.9210205078125, 247.32354468107224, 834.1371154785156, 302.19953916966915 ] ]
textvqa
train
null
null
null
what is that name on the bottom left?
de leon garnier
null
03e3bb402e4c2640.jpg
720
1,024
[ [ 274.24973487854004, 800.5401611328125, 353.29010009765625, 899.4457855224609 ], [ 27.00050264596939, 927.085693359375, 66.87358349561691, 975.5635070800781 ], [ 71.55273735523224, 916.8966674804688, 152.86957919597626, 985.1822891235352 ], [ 158.1322...
textvqa
train
null
null
null
Which two things that helps in removing CO2 and provide shoreline protection?
Mangroves, seagrass
null
38575.jpeg
3,000
2,349
[ [ 796.6084778308868, 593.0272167026997, 1384.9466890096664, 633.27939215675 ], [ 1468.1932926177979, 592.5170854926109, 2009.281486272812, 631.7906825188547 ], [ 469.2411571741104, 1260.426901459694, 800.2594336867332, 1299.8496849127114 ] ]
infographicsvqa
train
null
null
[ "Mangroves, seagrass" ]
What is the bag on, a couch or a bed?
bed
The bag is on a bed.
2322402.jpg
500
333
[ [ 386, 247, 498, 331 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "bag (3216237)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "furniture,on,o (3216238)" }, { "operation": "choose name", "dependencies": [ 1 ], "argument": "bed|couch" } ]
1. Identify the object in question; in this case, it is a bag that needs to be located within the image. 2. Once the bag is identified, observe what piece of furniture the bag is resting on. 3. The furniture could either be a couch or a bed, so closely examine characteristics in the image that distinguish a couch from ...
null
What animal is to the right of the bottle on the left?
horse
The animal is a horse.
2369633.jpg
500
375
[ [ 85, 0, 485, 375 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "bottle (2690951)" }, { "operation": "filter hposition", "dependencies": [ 0 ], "argument": "left" }, { "operation": "relate", "dependencies": [ 1 ], "argument": "animal,to the right of,s (213364...
1. Identify the bottle in the image. 2. Determine the position of the bottle relative to other objects, focusing on the bottle that is on the left. 3. Look to the right of the identified left-positioned bottle for an animal figure. 4. Determine the name of the animal that is situated to the right of the bottle on the l...
null
What is the animal to the right of the man on the left?
elephant
The animal is an elephant.
2389813.jpg
500
334
[ [ 82, 143, 201, 295 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "man (502377)" }, { "operation": "filter hposition", "dependencies": [ 0 ], "argument": "left" }, { "operation": "relate", "dependencies": [ 1 ], "argument": "animal,to the right of,s (502396)" ...
1. Identify the man in the image. 2. Determine the left side of the identified man from his own perspective. 3. Locate the animal that is positioned to the right of the identified man (from your perspective, this will be on the left side of the man). 4. Once the correct animal is located, determine the name of this ani...
null
What type of device is he holding?
wii controller
The man is holding the Wii controller.
2401910.jpg
375
500
[ [ 203, 367, 228, 394 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "he (1142794)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "device,holding,o (1142793)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the subject in the image referenced as 'he'. 2. Determine the object that the subject is holding. 3. Seek to identify the name or type of the device that the subject is holding.
null
what soda brand is shown on the x factor billboard?
pepsi
null
af64880695b72bd5.jpg
1,024
683
[ [ 216.0295867919922, 194.61027765274048, 256.72414779663086, 214.48867474123836 ] ]
textvqa
train
null
null
null
What do you think is the screen on?
game
The screen is on the game.
2415072.jpg
500
375
[ [ 109, 42, 409, 374 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "screen (146524)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,on,o (146523)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the object in question, which is the screen. 2. Determine the relationship the screen has with other elements in the image, focusing on the action or state it is involved in – in this case, what the screen is on. 3. Query or infer the specific nature of the content that the screen is displaying or the appli...
null
What is the guitar on?
wall
The guitar is on the wall.
2343614.jpg
364
500
[ [ 0, 0, 363, 260 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "guitar (921647)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,on,o (921638)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the guitar in the image. 2. Determine the relation that connects the guitar to another object or location in the image. 3. Utilize the relation identified to ascertain the name of the object or location the guitar is on.
null
Name the Indiviual or Corporate ?
charles burton
null
jpch0224_9.png
2,200
1,701
[ [ 217, 946, 1094, 1009 ], [ 217, 946, 1094, 1009 ] ]
docvqa
train
null
null
[ "charles burton", "Charles Burton" ]
what is the position of Dr. Ernest Reid ?
chmn.
null
yjpg0227_4.png
1,786
2,292
[ [ 700, 521, 1168, 555 ], [ 698, 957, 1119, 991 ], [ 628, 989, 1201, 1025 ], [ 683, 1193, 1156, 1227 ], [ 432, 1262, 539, 1292 ], [ 448, 1636, 535, 1666 ], [ 700, 521, 1168, 555 ]...
docvqa
train
null
null
[ "chmn.", "Chmn." ]
on what date is honorarium dated?
JUNE, 1976
null
njdv0228_12.png
1,714
711
[ [ 526, 214, 884, 245 ], [ 526, 214, 884, 245 ] ]
docvqa
train
null
null
[ "JUNE, 1976", "June, 1976" ]
Which category of people are at more risk for corona?
Older persons, those w/ Pre-existing Conditions
null
10576.jpeg
1,000
1,000
[ [ 664.105474948883, 622.5758194923401, 957.9703211784363, 642.4498856067657 ] ]
infographicsvqa
train
null
null
[ "Older persons, those w/ Pre-existing Conditions" ]
How many personal items are in this infographic?
4
null
10991.jpeg
1,000
3,233
[ [ 485.56041717529297, 1868.2550955414772, 510.10140776634216, 1899.219386935234 ] ]
infographicsvqa
train
null
null
[ "4" ]
Who is wearing a coat?
boy
The boy is wearing a coat.
2335497.jpg
500
332
[ [ 184, 21, 408, 333 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "coat (4626419)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "person,wearing,s (4626416)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the item referred to as a "coat" in the image. 2. Determine which person in the image is wearing the coat by noting the association between person and the item. 3. Once the person wearing the coat is identified, provide the name of that person.
null
what is the name of the pharmacy?
cvs
null
0c840974bf0844d9.jpg
1,024
600
[ [ 812.9953002929688, 79.99524772167206, 864.4015502929688, 98.55250306427479 ] ]
textvqa
train
null
null
null
how many think that someone could be hired within 90 seconds
more than 2000
null
30129.jpeg
1,200
4,056
[ [ 924.2448806762695, 352.03756749629974, 1152.3635923862457, 378.64716425165534 ] ]
infographicsvqa
train
null
null
[ "2000", "more than 2000" ]
What is the total population of prisoners in Unites states of America?
2,217,000
null
31693.jpeg
900
1,350
[ [ 215.42643159627914, 1054.6381741762161, 292.3511266708374, 1071.0434764623642 ] ]
infographicsvqa
train
null
null
[ "2217000", "2,217,000" ]
Who is the originator?
joe A. Leinster
null
jznv0001_1.png
1,692
2,245
[ [ 9, 593, 425, 644 ], [ 9, 593, 425, 644 ] ]
docvqa
train
null
null
[ "joe A. Leinster", "Joe A. Leinster" ]
what is the address of Ralston Purina Company?
835 South Eighth Street
null
qjfc0228_1.png
1,758
2,242
[ [ 390, 438, 793, 481 ] ]
docvqa
train
null
null
[ "835 South Eighth Street" ]
When is the Orange bowl match scheduled to be held?
December 31
null
32242.jpeg
800
2,400
[ [ 48.523110151290894, 1498.0324745178223, 137.3714804649353, 1511.1904453486204 ], [ 335.30852794647217, 1659.792423248291, 423.5679626464844, 1674.735975265503 ] ]
infographicsvqa
train
null
null
[ "December 31" ]
What's the cup on?
table
The cup is on the table.
2372722.jpg
500
375
[ [ 0, 202, 500, 375 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "cup (590582)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,on,o (590583)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the object in question from the image, which is the cup. 2. Determine the relationship between the cup and the surface it is on. 3. Investigate or enquire about the name of the object or surface that the cup is resting on.
null
who is the brewery?
brew dog
null
8e23c063fecb2667.jpg
1,024
683
[ [ 580.4153442382812, 157.6889992058277, 664.5166931152344, 193.51217703521252 ] ]
textvqa
train
null
null
null
What is located on top of the building the window is on the front of?
antenna
The antenna is on top of the building.
2376464.jpg
333
500
[ [ 280, 59, 312, 77 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "window (719963)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "building,on the front of,o (719962)" }, { "operation": "relate", "dependencies": [ 1 ], "argument": "_,on top of,...
1. Identify the window that is referenced in the question. 2. Determine which building has the window on its front. 3. Find out what is located on top of the identified building. 4. Request the name of the object that is on top of the building.
null
It is south of what?
Las Olas Boulevard
null
lmmf0227_4.png
870
1,888
[ [ 90, 1360, 739, 1392 ], [ 90, 1360, 739, 1392 ] ]
docvqa
train
null
null
[ "LAS OLAS BOULEVARD", "Las Olas Boulevard" ]
In the photographs, which cricketer is from West Indies
Brian Lara
null
41886.jpeg
900
7,118
[ [ 88.61189112067223, 236.4968486353755, 808.8889203965664, 352.07103241980076 ], [ 451.9297420978546, 520.8970253616571, 836.039936542511, 559.9900485724211 ] ]
infographicsvqa
train
null
null
[ "Brian Lara" ]
What percent of parents think that standardized tests are a positive thing?
44%
null
31591.jpeg
1,000
2,471
[ [ 894.2835330963135, 1148.92347240448, 967.6818251609802, 1182.6961447894573 ] ]
infographicsvqa
train
null
null
[ "44%" ]
Who is wearing the shirt?
man
The man is wearing a shirt.
2370464.jpg
500
331
[ [ 4, 305, 42, 331 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "shirt (2516840)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "person,wearing,s (2516839)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the item of clothing referred to as "shirt" in the image. 2. Determine the person who is associated with wearing the shirt. 3. Find out the name of the person who is wearing the shirt.
null
What's on the wall?
sign
The sign is on the wall.
3453.jpg
800
600
[ [ 597, 278, 650, 324 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "wall (4340053)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,on,s (4340052)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the wall within the image, which is the subject of the question. 2. Determine the relationship between the wall and any object that might be placed upon it. 3. Once the relationship of being "on" the wall is established for a specific object, ascertain the name of the object that is present on the wall.
null
Which district municipality is most densely populated?
Bojanala Platinum
null
35651.jpeg
890
942
[ [ 674.9692833423615, 183.01911395788193, 870.7018183171749, 200.57370779290795 ], [ 178.3015812933445, 589.4917094707489, 337.0840221643448, 603.2049257028848 ], [ 248.1173151731491, 825.8600317239761, 496.85579746961594, 841.462469112128 ] ]
infographicsvqa
train
null
null
[ "Bojanala Platinum" ]
What is the man that is to the left of the papers wearing?
ring
The man is wearing a ring.
2353907.jpg
375
500
[ [ 204, 406, 209, 410 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "papers (2049167)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "man,to the left of,s (2176718)" }, { "operation": "relate", "dependencies": [ 1 ], "argument": "_,wearing,o (234...
1. Identify the papers within the image. 2. Locate the man who is positioned to the left of the papers. 3. Examine what the man is wearing. 4. Report the name of the item the man is wearing.
null
which state of South Africa has been reported more number of confirmed cases - Limpopo or Northern Cape?
Limpopo
null
11502.jpeg
828
1,024
[ [ 562.928567647934, 359.2397155761719, 624.1309812068939, 371.6012268066406 ] ]
infographicsvqa
train
null
null
[ "Limpopo" ]
what is the name of the university mentioned in the given letter ?
university of California
null
fgyk0226_1.png
1,101
1,699
[ [ 88, 132, 559, 158 ], [ 88, 132, 559, 158 ] ]
docvqa
train
null
null
[ "university of california", "university of California" ]
What fruits are on the tree?
apples
The fruits are apples.
2406710.jpg
281
500
[ [ 135, 339, 200, 405 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "tree (291402)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "fruit,on,s (291401)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the tree within the image. 2. Locate the fruits that are on the tree. 3. Confirm the type of fruits by their visual characteristics and associate them with a commonly known fruit name.
null
What is the animal to the left of the young person?
elephant
The animal is an elephant.
2391382.jpg
500
334
[ [ 1, 2, 334, 334 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "person (1041035)" }, { "operation": "filter age", "dependencies": [ 0 ], "argument": "young" }, { "operation": "relate", "dependencies": [ 1 ], "argument": "animal,to the left of,s (1041043)" ...
1. Identify the person in the image. 2. Determine if the person identified is young. 3. Look for an animal that is located to the left of the young person. 4. Find the name of the animal that is positioned to the left of the young person.
null
Who is standing in the water of the river?
people
The people are standing in the water.
2376209.jpg
500
375
[ [ 289, 151, 360, 325 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "river (2378652)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "water,of,s (1915597)" }, { "operation": "relate", "dependencies": [ 1 ], "argument": "person,standing in,s (24149...
1. Identify the image of the river. 2. Locate the water within the image of the river. 3. Find any persons who are standing in the water in the image. 4. Confirm the name or identity of the individuals who are found standing in the river's water.
null
What is the date mentioned in the voucher?
June 18, 1998
null
nmfp0000_1.png
1,692
2,245
[ [ 292, 198, 462, 226 ], [ 1251, 267, 1422, 297 ] ]
docvqa
train
null
null
[ "June 18, 1998" ]
What is the person that is to the right of the man holding?
ball
The person is holding the ball.
2356220.jpg
500
350
[ [ 289, 224, 376, 314 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "man (3136967)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "person,to the right of,s (3095755)" }, { "operation": "relate", "dependencies": [ 1 ], "argument": "_,holding,o (22...
1. Identify the man in the image. 2. Locate the person who is to the right of the man. 3. Observe what the person to the right of the man is holding. 4. Determine the name of the object that the person is holding.
null
What is on the wall?
television
The TV is on the wall.
2345644.jpg
292
500
[ [ 56, 19, 219, 142 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "wall (2853050)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,on,s (2333147)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the feature referred to as "wall" in the image. 2. Determine the relationship indicating something being "on" in association with the identified wall. 3. Seek out the name or label of the object that has a relationship of being "on" the wall.
null
What is the average amount (in pounds) spent by women on home improvements?
1,382
null
37929.jpeg
1,500
1,910
[ [ 1197.0024704933167, 699.929426908493, 1439.859464764595, 721.7618350312114 ] ]
infographicsvqa
train
null
null
[ "1,382" ]
who wrote "beyond binary"?
brit mandelo
null
18080be3d85664f8.jpg
1,024
768
[ [ 515.2566528320312, 449.3394012451172, 587.2458724975586, 487.39065170288086 ] ]
textvqa
train
null
null
null
Mention the compensation for beck's AD position has been set per annum
$161,900
null
npbn0226_1.png
1,701
2,201
[ [ 251, 1243, 788, 1280 ] ]
docvqa
train
null
null
[ "$161,900" ]
what carrier is the phone from?
t-mobile
null
d62baa5f9596cb86.jpg
520
1,024
[ [ 77.32268214225769, 174.92543029785156, 151.77597880363464, 195.7226963043213 ] ]
textvqa
train
null
null
null
where is this gauge made?
japan
null
4f60ed496e41c410.jpg
1,024
774
[ [ 547.9765625, 158.09623011946678, 731.3827667236328, 209.9518439769745 ] ]
textvqa
train
null
null
null
who is the sponsor on the front of the jersey?
aeronautica
null
dc026b445ada7b61.jpg
1,024
640
[ [ 308.0213317871094, 354.62188720703125, 542.0754699707031, 411.7850923538208 ] ]
textvqa
train
null
null
null
Where is the man to the right of the person standing on?
street
The man is standing on the street.
2330203.jpg
500
326
[ [ 0, 234, 61, 325 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "person (4667235)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "man,to the right of,s (4667191)" }, { "operation": "relate", "dependencies": [ 1 ], "argument": "place,standing ...
1. Identify the person mentioned in the question within the visual context provided. 2. Locate the man who is to the right of the identified person. 3. Determine the place where the identified man is standing on. 4. Provide the name or description of that place.
null
what is rain and cooler weather always reduce the risk of bushfires
myth
null
34049.jpeg
724
1,024
[ [ 189.8763406276703, 152.64007568359375, 535.5868092775345, 188.03141021728516 ], [ 82.52261698246002, 361.2158203125, 121.328584253788, 374.71707820892334 ], [ 82.90638339519501, 563.50732421875, 120.61546562612057, 577.0019845962524 ] ]
infographicsvqa
train
null
null
[ "myth" ]
Which type of digital advertising contributes to the major part of revenue in the online newspaper industry?
DISPLAY ADS
null
43630.jpeg
2,000
1,200
[ [ 49.007415771484375, 333.0054759979248, 318.3113932609558, 350.939441844821 ], [ 847.6755023002625, 403.5125970840454, 999.5507448911667, 417.7598610520363 ], [ 368.20974946022034, 522.9092001914978, 501.27603113651276, 539.5488303154707 ], [ 847.6755...
infographicsvqa
train
null
null
[ "DISPLAY ADS", "DISPLAY OR BANNER ADS" ]
What is lying in the grass?
frisbee
The frisbee is lying in the grass.
2343330.jpg
500
334
[ [ 53, 62, 270, 279 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "grass (2987267)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,lying in,s (2987266)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the area of grass in the image. 2. Look for an object within the grass that is positioned as if it is lying down or resting on the grass. 3. Determine the name of the object that is found lying in the grass.
null
what is the venue for meeting ?
Dean's office
null
zshd0227_1.png
1,783
1,181
[ [ 597, 654, 821, 683 ], [ 597, 654, 821, 683 ] ]
docvqa
train
null
null
[ "DEAN'S OFFICE", "Dean's office" ]
what is the date of issue?
18 Mar 74
null
rkdv0228_15.png
1,516
689
[ [ 688, 120, 865, 154 ], [ 688, 120, 865, 154 ] ]
docvqa
train
null
null
[ "18 mar 74", "18 Mar 74" ]
What's the man on?
skateboard
The man is on the skateboard.
2355050.jpg
333
500
[ [ 107, 351, 172, 380 ] ]
gqa
train
[ { "operation": "select", "dependencies": [], "argument": "man (2182151)" }, { "operation": "relate", "dependencies": [ 0 ], "argument": "_,on,o (3570677)" }, { "operation": "query", "dependencies": [ 1 ], "argument": "name" } ]
1. Identify the subject in the image, which is the 'man'. 2. Determine the relationship that describes the position or action of the man within the image. The specific relationship of interest is 'on', indicating the man is positioned on top of something. 3. Query the object that the man is 'on', referring to the item ...
null
How many coffee cups are in this infographic?
1
null
42139.jpeg
600
1,541
[ [ 84.2801034450531, 283.14417420327663, 93.99058520793915, 302.2015530914068 ] ]
infographicsvqa
train
null
null
[ "1" ]
End of preview.

Vision-RL² training data

Paper: https://arxiv.org/abs/2609.19745

Training data for Vision-RL²: Region-Level Policy Optimization for Fine-grained MLLM Perception — the stage-1 (SD-RPN self-distilled pseudo-label) response corpora and the stage-2 (region-level RL) pools with their cached evidence maps, for all four backbones: Qwen3.5-4B, Qwen3.5-9B, Qwen2.5-VL-7B and Gemma-4-12B-it.

Code: https://github.com/YuHengsss/VisionRL2. Images: YuhengSSS/RoITraining.

Contents

sdrpn_corpora/                          stage 1 — self-generated response corpora
  qwen3_5_4b_response_corpus.jsonl        49,498 rows
  qwen3_5_9b_response_corpus.jsonl        49,999 rows
  gemma4_12b_response_corpus.jsonl        49,498 rows (tier 560)
rl_pools/                               stage 2 — RL pools (7,000 rows each)
  candidates_visualcot_50k.jsonl          50,000 candidate QAs the pools are drawn from
  rl_pool_qwen3_5_4b.jsonl
  rl_pool_qwen3_5_9b.jsonl
  rl_pool_qwen2_5_vl_7b.jsonl
  rl_pool_gemma4_12b.jsonl
ev_maps/                                cached response→image evidence maps (7,000 .pt each)
  ev_maps_qwen3_5_4b.tar
  ev_maps_qwen3_5_9b.tar
  ev_maps_qwen2_5_vl_7b.tar
  ev_maps_gemma4_12b.tar

Download and layout

hf download YuhengSSS/VisionRL2-data --repo-type dataset --local-dir data/VisionRL2-data
mkdir -p data/ev_maps
for t in data/VisionRL2-data/ev_maps/*.tar; do tar -xf "$t" -C data/ev_maps; done

giving data/ev_maps/ev_maps_qwen3_5_4b/00000012.pt, and so on.

Images come from YuhengSSS/RoITraining (tars: gqa, textvqa, spdocvqaDocVQA, infographicsvqa). Extract them so that DATASET_ROOT looks like:

DATASET_ROOT/
  gqa/images/<image>
  textvqa/train_images/<image>
  DocVQA/<image>
  infographicsvqa/infographicsvqa_images/<image>

DATASET_ROOT defaults to datasets; individual roots can be overridden with DS_IMAGE_ROOTS="gqa=/abs/path,docvqa=/abs/path,...".

Regenerating these files

Every file here can be rebuilt from the candidates plus your own checkpoints, using the release repo:

target command
Qwen3.5 corpora data_prep/build_corpus_qwen3_5.sh (split → generate v1/v2 → merge)
Gemma-4 corpus scripts/train_sdrpn_gemma4.sh stages gen + merge
Qwen3.5 pools data_prep/build_pool_qwen3_5.sh (filter → compose → evidence → maps)
Qwen2.5-VL pool data_prep/build_pool_qwen2_5_vl.sh
Gemma-4 pool data_prep/build_pool_gemma4.sh

Row schemas

SD-RPN corpus row (sdrpn_corpora/)

One row per QA; the response is generated by the frozen backbone itself, and the pseudo-label is the response-to-image attention of that response.

field meaning
dataset source tag: gqa / textvqa / docvqa / infographicsvqa — selects the image root
image image path relative to that dataset's image root
question raw question
prompted_question the exact prompt that was fed to the backbone (prompt style baked in)
answer gold short answer
response the backbone's own response (the pseudo-label is read off its attention)

The prompted_question of a row fixes its prompt style: gqa rows carry the bounding-box task suffix and textvqa rows the "single word or phrase" suffix (style v1, square-padded images, up to 1,024 visual tokens), while docvqa and infographicsvqa rows carry the [Visual Evidence] … [Answer] prompt (style v2, no square padding, up to 576 visual tokens). Both passes decode greedily with 512 new tokens; rows whose response came back empty were dropped.

The Qwen corpora carry the extra VisualCoT provenance fields (width, height, bboxs, split, full_answer, reasoning, thought, possible_answers); the trainer ignores them. The pseudo-label style per row (v1 = mean-over-response-tokens map, v2 = single-region peak-ratio union) is derived from the dataset tag via DS_TO_LABEL_VERSION in qwen-vl-finetune/qwenvl/data/__init__.py (gqa/textvqa → v1, docvqa/infographicsvqa → v2).

The Gemma corpus instead carries the style as an explicit per-row field and uses full dataset-root-relative image paths:

field meaning
dataset, question, answer, response as above
image path relative to DATASET_ROOT (e.g. gqa/images/2337160.jpg)
question_id, src_row provenance into candidates_visualcot_50k.jsonl
version v1 (gqa + textvqa) or v2 (docvqa + infographicsvqa) — consumed by qwen_src/gemma4_unified/data_gemma_stage1.py

RL pool row (rl_pools/rl_pool_*.jsonl)

field meaning
sample_id index into candidates_visualcot_50k.jsonl
dataset source tag (image root selector)
image image path relative to that dataset's image root
question question (the eval answer suffix is appended by the loader)
gold_answer gold answer scored by the frozen reader
feat_hw [H, W] of the SD-RPN heatmap grid at pool-build resolution
branch K==1 / K>=2 — number of connected components in the stage-1 proposal
K, K_topR components found / components kept as actions
n_actions size of the enumerated action set
reward_mean, reward_std stage-1 reward statistics over that action set (the ranking signal)
ev_maps_path ev_maps_<backbone>/<sample_id>.pt — the cached evidence maps for this row

candidates_visualcot_50k.jsonl is the raw 50k VisualCoT candidate set the pools are drawn from (gqa 20k, textvqa 10k, docvqa 10k, infographicsvqa 10k).

EV_MAPS_ROOT convention

ev_maps_path is relative. The loader (qwen-vl-finetune/qwenvl/train/region_level_grpo/dataset.py) resolves it in order: as given, then $EV_MAPS_ROOT/<ev_maps_path>, then $EV_MAPS_ROOT/<basename>, then the pool jsonl's own directory. With the layout above, set:

export EV_MAPS_ROOT=data/ev_maps

Each .pt holds {"maps": uint8 tensor (n_layers, Hg, Wg)} — the frozen backbone's response-to-image attention at the probed layers, binarised per layer. They feed the additive (source-map) group of the RL objective, which recovers evidence the policy never proposed. A row that declares an ev_maps_path whose file cannot be found is a hard error (a silent miss would turn that group off).

How the 7k pools were selected

The pool is built with the backbone's own SD-RPN (stage-1) checkpoint:

  1. Run the backbone's SD-RPN over the 50k candidates; drop gqa and chartqa, and drop rows whose gold region covers more than 10% of the image.
  2. Enumerate the action set (top R = 6 components: the empty action plus singleton drops), score every action with the frozen reader, and keep the top 20% of rows by per-sample reward standard deviation — rows where the choice of region actually moves the reward.
  3. Compose the final pool: 5,000 InfographicVQA + 1,000 TextVQA + 1,000 DocVQA = 7,000.
  4. Cache the response-to-image evidence maps for every kept row (ev_maps/), from the backbone's own evidence responses.

The heatmap gate at step 1 differs per family (Qwen3.5: fixed threshold 0.02; Qwen2.5-VL-7B and Gemma-4: peak-ratio).

Provenance note. rl_pool_qwen3_5_4b, rl_pool_qwen3_5_9b and rl_pool_gemma4_12b were each selected by their own SD-RPN and are therefore different row sets (the 4B and 9B pools share 2,780 of 7,000 rows). rl_pool_qwen2_5_vl_7b is the exception: it reuses the Qwen3.5-4B row selection verbatim (identical 7,000 sample_ids and identical step-1/2 statistics, reordered) and only steps 3–4 are its own — the evidence responses and cached attention maps in ev_maps_qwen2_5_vl_7b.tar come from Qwen2.5-VL-7B itself. That is what the paper's 7B run trained on.

The release repo ships the full pipeline (data_prep/build_pool_qwen3_5.sh, build_pool_qwen2_5_vl.sh, build_pool_gemma4.sh) so the pools can be regenerated on your own checkpoint.

License and provenance

The QAs and images are derived from Visual-CoT and, through it, from GQA, TextVQA, DocVQA (SP-DocVQA) and InfographicVQA. All original licenses and terms of those datasets apply; this release adds only model-generated responses, region statistics and cached attention maps, and is intended for research use. No images are redistributed here — only jsonl rows and .pt caches that reference them by relative path.

Citation

@article{shi2026visionrl2,
  title   = {Region-Level Policy Optimization for Fine-grained MLLM Perception},
  author  = {Shi, Yuheng and Pei, Xiaohuan and Dong, Minjing and Xu, Chang},
  journal = {arXiv preprint arXiv:2609.19745},
  year    = {2026}
}

@inproceedings{shi2026sdrpn,
  title     = {Catching the Details: Self-Distilled RoI Predictors for Fine-Grained MLLM Perception},
  author    = {Shi, Yuheng and Pei, Xiaohuan and Dong, Minjing and Xu, Chang},
  booktitle = {ICLR},
  year      = {2026}
}
Downloads last month
28

Collection including YuhengSSS/VisionRL2-data

Paper for YuhengSSS/VisionRL2-data