row_uid stringlengths 16 16 | source stringclasses 1
value | task stringclasses 1
value | question stringlengths 234 631 | options stringclasses 0
values | answer stringclasses 22
values | num_input_images int32 1 1 | num_steps int32 2 5 | input_image_0 imagewidth (px) 512 512 | thought_0 stringlengths 360 816 | target_image_0 imagewidth (px) 384 911 | thought_1 stringlengths 311 723 | target_image_1 imagewidth (px) 384 911 | thought_2 stringlengths 278 625 ⌀ | target_image_2 imagewidth (px) 384 911 ⌀ | thought_3 stringlengths 241 619 ⌀ | target_image_3 imagewidth (px) 384 911 ⌀ | thought_4 stringlengths 217 604 ⌀ | target_image_4 imagewidth (px) 384 384 ⌀ | thought_5 stringlengths 180 564 | split stringclasses 1
value | license stringclasses 1
value | source_scene_corpus stringclasses 1
value | est_tokens int32 12.1k 24.5k | filter_flags stringclasses 8
values |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
53e29d83d9b16f6f | zebra_multihop | multihop_counting | Given the initial arrangement of objects displayed in the image below, which is viewed from a Right0 angle, you are tasked with performing two specific operations. First, add two ORANGE PYRAMID objects. Second, remove two GREEN objects. After these operations, what is the total number of objects present? | null | 8 | 1 | 3 | I need the final object count after two operations: add two orange pyramids, then remove two green objects. In the given right-side view I can identify an orange cube, a green sphere, a blue pyramid, a large red cone, a green capsule, a red torus, and an orange capsule. But the camera is low and the red cone sits in fr... | The top-down map settles the initial scene: besides the objects I listed, a small blue torus sits between the orange cube and the green capsule, exactly where the red cone blocked the low view. So the scene starts with eight objects: cube, sphere, blue pyramid, blue torus, red cone, green capsule, red torus, orange cap... | This picture shows the map after the addition: the original eight objects are all still in their places, and two orange pyramids now stand on the open ground at the lower right, making ten objects in total. The second operation removes two green objects. The only green objects on the map are the small green sphere at t... | null | Not supported with pagination yet | null | Not supported with pagination yet | In the final map the two green objects are gone. Counting what remains: the blue pyramid, the orange cube, the blue torus, the red torus, the orange capsule, the red cone, and the two orange pyramids. That is six single objects plus the two pyramids, eight objects in total. So after adding two orange pyramids and remov... | train | cc-by-nc-4.0 | procedural_grid | 16,813 | S5.replay_unsupported | ||||
2340a62839ff69a4 | zebra_multihop | multihop_counting | Given an initial arrangement of objects as shown in the image below, viewed from a Back45 angle, perform a series of operations. First, add 2 purple capsule objects. Second, remove zero black objects. Third, remove all purple objects. After these steps, determine the total count of blue cylinders remaining. | null | 2 | 1 | 3 | The question asks for the blue cylinders remaining after adding 2 purple capsules, removing zero black objects, and removing all purple objects. The input view shows 2 blue cylinders (a light one at left, a dark one at right), a small blue cube, a black cube, an orange dodecahedron, 1 purple capsule, and 1 purple spher... | In this picture the 2 added capsules are present: one at the top and one to the right of the dark blue cylinder, so with the original capsule there are now 3 purple capsules plus the purple sphere. The blue objects are unchanged: 2 blue cylinders and the small blue cube. The next operation removes zero black objects, s... | The second picture matches the first exactly, as expected: 3 purple capsules, 1 purple sphere, 2 blue cylinders, 1 blue cube, 1 black cube, and 1 orange dodecahedron. The final operation removes every purple object, so all four purple pieces disappear and nothing else does. I will sketch the same viewpoint with the pur... | null | Not supported with pagination yet | null | Not supported with pagination yet | In the last picture every purple object is gone. What remains is the light blue cylinder at left, the dark blue cylinder at right, the small blue cube, the black cube, and the orange dodecahedron. Counting blue cylinders only: the light cylinder and the dark cylinder, two in all; the blue cube is a cube, not a cylinder... | train | cc-by-nc-4.0 | procedural_grid | 16,849 | S5.replay_unsupported,S8.phash_near_but_distinct | ||||
5c4b5ad058c95d34 | zebra_multihop | multihop_counting | Given an initial arrangement of objects as depicted in the image below, observed from a Front0 perspective, you are required to perform a sequence of operations:
Step 1: Introduce 4 new BLUE PYRAMID objects into the scene.
Step 2: Eliminate 1 PURPLE object from the collection.
Step 3: Remove 2 BLUE objects.
Step 4: Do ... | null | 1 | 1 | 5 | I must end with the number of cube-shaped objects left after a series of additions and removals. In the front view I can pick out one gray cube, a small purple ring resting on top of it, a tall magenta capsule nearby, one blue pyramid, one blue sphere, and a single orange sphere. The low angle is not safe for tracking:... | The high view settles the layout: exactly one gray cube, with the purple ring on top of it, the magenta capsule standing clear beside it, the blue pyramid and blue sphere out to one side, and the orange sphere off to the other. No second cube is hidden anywhere in the cluster. The first operation adds four blue pyramid... | This picture shows the four new blue pyramids spread around the scene in addition to the original one, and the gray cube with its purple ring is still sitting in the middle of the group. So the set is now fixed: one cube, a purple ring, the magenta capsule, five blue pyramids, a blue sphere, and the orange ball. The ne... | The capsule is gone from the picture; the purple ring on the cube, the five blue pyramids, the blue sphere and the orange ball are all still in their places, and the cube itself is untouched. The next operation removes two blue objects, so I will choose the blue sphere and one of the added pyramids at the far end of th... | In this view the blue sphere and that far pyramid have disappeared; the remaining blue pyramids, the gray cube with its purple ring, and the orange ball are all still present in the same spots. The last operation says no orange object is removed, so nothing changes at all. I will redraw the arrangement exactly as it is... | The final view shows the gray cube with its purple ring still sitting on top, the remaining blue pyramids, and the orange ball, with nothing else in the scene. Scanning the shapes: the pyramids are four-sided points, the ball is round, the ring is a loop, and the only object with six square faces is the single gray cub... | train | cc-by-nc-4.0 | procedural_grid | 24,494 | S5.replay_unsupported,S9.px=384 | ||||||
9c92248c96d8c96b | zebra_multihop | multihop_counting | Given an initial arrangement of objects as seen from a Front0 angle, you are required to perform a series of modifications:
Step 1: Introduce 4 new PURPLE PYRAMID objects.
Step 2: Add 2 RED CUBE objects.
Step 3: Eliminate 1 DONUT object from the scene.
Step 4: Remove all PYRAMID objects.
After completing all these step... | null | 2 | 1 | 5 | The question asks for the number of red objects left after adding 4 purple pyramids, adding 2 red cubes, removing 1 donut, and then removing all pyramids. The front view shows a black cube, a brown ring, an orange ring, a red pyramid, and a black cone, and the pyramid partly hides the orange ring, so I cannot track the... | On the map I see exactly the five starting objects: the dark cube at left, the brown ring and the orange ring in the middle, the red square that is the pyramid seen from above, and the dark half-disk cone at right. That confirms my starting inventory and gives me stable positions to work on. The first change adds four ... | The map now carries four magenta shapes to the right of the rings on top of the original five marks, and nothing else moved, so the purple pyramids are in. The next change adds two red cubes, which from above read as red square-ish blocks. I will mark two red shapes at free places, one at the top right and one just bel... | I can verify the additions: two red blocks sit on the map, one high at the right and one just below the orange ring, alongside the four magenta shapes, the red square, the two rings, the cube and the cone. The next change removes one donut, so I will erase one of the two rings. I choose the orange ring, the one partly ... | The orange ring is gone and only the brown ring remains, so one donut is out. The last change removes every pyramid: that is the red square marking the original pyramid, plus the four magenta shapes. I will clear those five flat shapes and keep the dark cube, the brown ring, the two red blocks, and the dark half-disk c... | The final map holds only four objects: the dark cube, the brown ring, the two red blocks, and the dark half-disk cone. Counting the red ones item by item, the two red cube blocks are red, and the red square of the original pyramid has been cleared, so no other red object remains. That makes 2 red objects. The answer is... | train | cc-by-nc-4.0 | procedural_grid | 24,250 | S5.replay_unsupported,S9.px=384 | ||||||
319f1f03e6e09404 | zebra_multihop | multihop_counting | Given the initial set of objects shown in the image below, observed from a Front0 perspective, execute the following two sequential operations: First, eliminate all objects identified as 'DONUT'. Second, remove one 'ORANGE' object. After completing these steps, determine the total count of 'cube' objects remaining. | null | 4 | 1 | 3 | I must count the cubes left after deleting every donut and one orange object. The front view shows a brown cube on the left, a blue pyramid, a low purple hexagon, a small brown solid behind the tall purple cube, that cube, a purple cube in front, a red cube, and a red ring on the right. At this low angle the small soli... | The overhead map settles the shapes: on the left a brown cube; a blue pyramid; a flat purple hexagon; a small brown solid; a tall purple cube at the right; a purple cube in front of it; a red cube beside that; and a red ring-shaped donut at the lower right. That is four cubes in total so far. The first operation delete... | In this picture the red ring is gone and everything else stands as before: brown cube, blue pyramid, flat purple hexagon, small brown solid, tall purple cube, front purple cube, red cube. Now I remove one orange object. The flat cube on the left reads as brown, while the small solid up top reads as orange, so the objec... | null | Not supported with pagination yet | null | Not supported with pagination yet | The last picture shows the blue pyramid, the flat purple hexagon, the brown cube on the left, the tall purple cube, the front purple cube, and the red cube; the donut and the small solid are both absent. Counting the cubes one by one: the brown cube, the tall purple cube, the front purple cube, and the red cube, which ... | train | cc-by-nc-4.0 | procedural_grid | 16,792 | S5.replay_unsupported | ||||
d4e5d594996fa914 | zebra_multihop | multihop_counting | Given an initial set of objects displayed in the image below, observed from a Right0 degree angle, we need to perform two sequential operations: First, eliminate all orange-colored objects. Second, introduce three new blue cube objects into the scene. Finally, determine the total count of red objects remaining after th... | null | 1 | 1 | 3 | The question asks me to remove all orange objects, then add three blue cubes, and finally count the red objects. From this low right-side angle I see a purple cube, an orange pyramid partly hidden behind it, a brown pyramid in front, and a red capsule on the right. The horizon and the overlap make the full set hard to ... | This high view lays the set on open sand: red capsule on the left, a small brown pyramid and the purple cube in the center, an orange ball and a larger rust-faced pyramid on the right, and a purple ring at the bottom right. The map shows the orange piece as a ball rather than a pyramid, but I read it as it stands: the ... | The redrawn view keeps the red capsule, the purple cube, and the purple ring on the same ground. The orange ball and the rust-faced pyramid are gone as planned, but the small brown pyramid has also disappeared; I take the picture as the remaining set, and since that pyramid was not red, nothing I track changes. What is... | null | Not supported with pagination yet | null | Not supported with pagination yet | This final view shows the red capsule on the left, the purple cube and the purple ring on the right, and three blue cubes on the sand in the foreground: one to the left, two in the center. Counting the red objects among these, only the capsule is red; the cube, the ring, and the three added cubes are purple or blue. So... | train | cc-by-nc-4.0 | procedural_grid | 16,826 | S5.replay_unsupported | ||||
cebda3352aa674ea | zebra_multihop | multihop_counting | Given an initial arrangement of objects, as depicted in the front view, you are required to perform a series of transformations. The operations are as follows:
Step 1: Eliminate one DONUT object from the scene.
Step 2: Eliminate zero SPHERE objects.
Step 3: Introduce four BLUE CUBE objects into the scene.
Step 4: Remov... | null | 2 | 1 | 5 | I need the final number of orange objects after four changes. The front view shows a black donut, a red cone, a red sphere, an orange cylinder, a blue cone, and an orange cone on the ground. I cannot keep all removals and additions in my head from this angle, especially since later cubes may crowd the view. So I sketch... | On this map the black donut is gone, leaving exactly five objects: the red cone at the right, the red sphere in front, the orange cylinder in the middle, the orange cone behind it, and the blue cone at the left. The next instruction removes zero spheres, so the red sphere must stay and nothing else changes. I therefore... | This redrawn map still shows the same five objects in the same places, with no sphere missing and no cube present. That settles the state after the sphere step, but the cube step has not been applied yet. I need a base that clearly has empty ground where new cubes can be placed, so I keep this exact arrangement in view... | The base map has the red cone, red sphere, orange cylinder, orange cone, and blue cone. Now I introduce four blue cubes, placing one near the upper right, one to the middle right, and two close together at the front bottom, each on an empty part of the ground. The red objects remain for now because their removal comes ... | This picture shows the four blue cubes together with the red cone and red sphere still present, plus the orange cylinder, orange cone, and blue cone. The final instruction removes every red object, so I erase the red cone on the right and the red sphere in front. The remaining objects are the four blue cubes, the blue ... | In the final map I count orange objects one by one. The orange cylinder stands in the middle, and the orange cone stands just behind it. The other visible objects are four blue cubes and one blue cone; no red, black, or other orange items remain. Therefore the total number of orange-colored objects is 2. | train | cc-by-nc-4.0 | procedural_grid | 24,231 | S5.replay_unsupported,S9.px=384 | ||||||
7ac4aa0410787f57 | zebra_multihop | multihop_counting | Given an initial setup of various 3D objects, as seen from the Back0 angle in the image below, the task is to perform a sequence of removal operations. First, eliminate all black-colored objects. Next, remove all objects identified as cones. Finally, discard all dodecahedron-shaped objects. After these steps are comple... | null | 1 | 1 | 4 | The question asks me to remove every black object, then every cone, then every dodecahedron, and finally count the blue objects left. The view I am given is low, and the dark object just behind the blue capsule is almost hidden by it, so I cannot yet be sure of the full list, nor whether the cluster on the right hides ... | The map I sketch shows six objects, each clear of the others: the blue capsule at the lower left, the black cylinder at the upper left, the blue cone at the top middle, the green sphere in the middle, the small brown dodecahedron at the upper right, and the black dodecahedron just to the right of the sphere. The first ... | In this picture the two black objects are gone; four remain: the blue capsule at the left, the blue cone at the top, the green sphere in the middle, and the brown dodecahedron at the upper right. That settles the first step. The next operation removes all cones, so the blue cone must go. I will draw the scene with the ... | Now three objects stand in the picture: the blue capsule at the left, the green sphere in the middle, and the brown dodecahedron at the upper right; the cone is absent. The last operation discards every dodecahedron-shaped object, which removes the brown one. I will draw the scene with the dodecahedron gone, leaving on... | null | Not supported with pagination yet | The last picture shows exactly two objects: the blue capsule and the green sphere. Of these, only the capsule is blue-colored; the other object is green. So the number of blue objects remaining after all the removals is 1. | train | cc-by-nc-4.0 | procedural_grid | 21,283 | S5.replay_unsupported | |||||
46d25a05f4f42265 | zebra_multihop | multihop_counting | Given the initial arrangement of objects shown below, viewed from the Front0 angle, you need to perform a sequence of operations. First, add three new black pyramid objects to the scene. Second, remove all objects that are pyramids, regardless of their color. After these operations, how many black pyramid objects remai... | null | 0 | 1 | 3 | I need the final count of black pyramids after adding three black pyramids and then removing every pyramid. The front view shows a red cone, a red sphere, a green ring, an orange sphere, a blue cone, a black pyramid on the right, two dark spheres, and a magenta pointed object partly hidden behind the blue cone. The mag... | The top-down map lays the scene flat. I can now read the pointed objects clearly: the magenta one sits just behind the blue cone and is a pyramid, and the black one to the right is a pyramid seen from above. No black pyramid is hidden elsewhere, so the start has exactly one black pyramid and one magenta pyramid. The fi... | This picture now shows the added pyramids: the original black pyramid on the right plus three new black pyramids near the ring, near the red sphere, and at the bottom, making four black pyramids, alongside the magenta pyramid. That settles the state after the first operation. What is still open is the second operation,... | null | Not supported with pagination yet | null | Not supported with pagination yet | In this final top-down view, none of the four black pyramids is present; the right-side black pyramid and the three I added have all gone, and only the cone, the spheres, and the ring remain. The magenta pointed object still shows, but the question concerns only black pyramids, and the removal step has taken out every ... | train | cc-by-nc-4.0 | procedural_grid | 16,892 | S5.replay_unsupported | ||||
495e3b6b298d48c3 | zebra_multihop | multihop_counting | Given an initial arrangement of objects viewed from a top-down perspective, as shown in the image, perform two sequential operations: first, add 4 blue capsule objects, and then remove 1 red object. After these operations, determine the total number of capsule objects remaining. | null | 7 | 1 | 3 | The question asks for the number of capsules left after two operations: adding 4 blue capsules, then removing 1 red object. The top-down input shows nine objects, but straight from above, capsules, spheres, and cones all flatten into round blobs, so I cannot yet tell which of the red, green, and blue pieces are capsule... | In the slanted view the shapes are now distinct: a red capsule, a green capsule, and a blue capsule stand among a red sphere, three tori (red, blue, magenta), a green cone, and a black cone. That is 3 capsules initially. The first operation adds 4 blue capsules, so the next picture I will draw is this same scene with f... | This picture shows the original blue capsule now accompanied by four new blue capsules, so there are 5 blue capsules; together with the red capsule and the green capsule that makes 7 capsules. But the second operation removes one red object, and I do not yet know whether that red piece is a capsule or one of the round ... | null | Not supported with pagination yet | null | Not supported with pagination yet | In this picture the red capsule is still standing in the middle of the scene; the red object that was removed was one of the round red pieces, not a capsule. So the capsule count is unchanged from the previous view: 5 blue capsules, the 1 red capsule, and the 1 green capsule. Counting them item by item gives 5 + 1 + 1 ... | train | cc-by-nc-4.0 | procedural_grid | 16,742 | S5.replay_unsupported | ||||
eb19197e2adf9727 | zebra_multihop | multihop_counting | Given the initial arrangement of objects shown in the image below, which is viewed from the Right0 angle, perform a series of removal operations. First, remove one SPHERE object. Next, eliminate all PURPLE objects. Finally, remove all GREEN objects. After these steps, determine the total count of orange objects remaini... | null | 2 | 1 | 4 | I need the final number of orange objects after three removals. The input is a low, ground-level view: a red cube at left, an orange pyramid partly behind a green polyhedron, a brown metal sphere in the middle, a magenta pyramid and a flat magenta torus at right, and an orange rubber sphere in front right. At this angl... | This top-down map shows all seven bodies separated: red cube at lower center, orange pyramid lower left, brown metal sphere mid-left, green polyhedron mid-right, magenta pyramid upper mid-left, flat magenta torus center, orange rubber sphere upper right. The first operation removes one sphere. There are exactly two sph... | With the brown sphere gone, this map holds six: red cube, orange pyramid, orange sphere, green polyhedron, magenta pyramid, magenta torus. The next operation removes every purple object. I see two magenta bodies, the small pyramid up top and the flat torus at the center; both count as purple, so both leave. I will redr... | Now only four bodies remain in this map: red cube lower center, orange pyramid lower left, orange sphere upper right, and the green polyhedron mid-right. The last operation removes all green objects, and the green polyhedron is the only green one, so it goes. I will redraw the map with the green polyhedron omitted, lea... | null | Not supported with pagination yet | This final map shows three objects: the red cube at lower center, the orange pyramid at lower left, and the orange sphere at upper right. Counting orange bodies one by one, the pyramid is orange and the sphere is orange, while the cube is red and not orange. That gives two orange objects remaining, so the answer is 2. | train | cc-by-nc-4.0 | procedural_grid | 21,341 | S5.replay_unsupported | |||||
877dc4735165491b | zebra_multihop | multihop_counting | Analyze the initial scene which shows several 3D objects viewed from the Back0 angle. Perform a series of removal operations: first, remove 0 orange objects; second, remove 2 capsule objects; third, remove all cylinder objects; and finally, remove 1 red object. After all operations are complete, determine the count of ... | null | 1 | 1 | 5 | I must count green spheres after four removals, so I need to track every object through the changes. The input view shows a red capsule at the front left, an orange cylinder just behind it, a small red octahedron in the middle, a magenta cone in the center, a small green sphere tucked behind the cone, and a blue capsul... | On my map, six objects stand: the red capsule at the left, the orange cylinder beside it, the red octahedron in the middle, the magenta cone below it, the green sphere to the right of the cone, and the blue capsule at the right. The first operation removes 0 orange objects. The orange cylinder is the only orange object... | This picture confirms the no-op: all six objects are still present, including both capsules, the red one at the left and the blue one at the right. The next operation removes 2 capsule objects, and I have exactly two capsules, so both must go. I will draw the map again with the red and blue capsules erased, leaving the... | The map now shows four objects: the orange cylinder at the top, the red octahedron in the middle, the magenta cone at the left, and the green sphere at the right. The capsules are gone as expected. The next operation removes all cylinder objects. Only the orange object is a cylinder, so I will draw the map without it, ... | This picture shows three objects: the red octahedron in the middle, the magenta cone at the left, and the green sphere at the right. The last operation removes 1 red object. The only red object left is the octahedron, since the red capsule was already removed, so I will draw the map with the octahedron erased, leaving ... | The final map shows exactly two objects: the magenta cone and a single green sphere. The green sphere was never targeted: the removals covered orange objects (zero of them), the two capsules, the one cylinder, and the one red object, none of which was the green sphere. So the count of green spheres remaining is 1. | train | cc-by-nc-4.0 | procedural_grid | 24,224 | S5.replay_unsupported,S9.px=384 | ||||||
1d9f7b25d4f1ac96 | zebra_multihop | multihop_counting | Consider the initial arrangement of objects shown in the image, viewed from the Back0 angle. Your task is to apply a series of transformations: first, add two blue donut shapes; next, eliminate all sphere objects; and finally, remove all pyramid objects. After these operations, determine the total count of orange objec... | null | 1 | 1 | 4 | I need the number of orange objects after three operations: adding two blue donuts, removing all spheres, removing all pyramids. The input image shows a red pyramid front left, an orange cylinder, a purple cylinder, a small dark shape partly hidden behind the purple cylinder, a red sphere, and a green cylinder. From th... | From this higher Back0 view I can resolve the scene: a red pyramid front left, an orange cylinder in the center, a purple cylinder, a dark olive object behind the purple cylinder, a red sphere on the right, and a green cylinder front right. The dark object tapers smoothly to a point, so it reads as a cone, not a facete... | This picture shows the two blue donuts in place: one just left of the purple cylinder, the other right of the dark cone. All original objects are still present: the red pyramid front left, the orange cylinder, the purple cylinder, the dark cone, the red sphere, and the green cylinder. That settles the add-donuts step. ... | In this picture the red sphere is gone, leaving the red pyramid, the orange cylinder, the purple cylinder, the dark cone, the two blue donuts, and the green cylinder. The final operation removes every pyramid. The red object front left has flat triangular faces meeting at an apex, so it is a pyramid and must go. The da... | null | Not supported with pagination yet | The last picture shows what remains: the orange cylinder in the center, the purple cylinder, the dark cone behind it, the two blue donuts, and the green cylinder. The red pyramid and the red sphere are both gone. Counting orange objects item by item, the orange cylinder is the only orange one; the purple, green, blue, ... | train | cc-by-nc-4.0 | procedural_grid | 21,437 | S5.replay_unsupported | |||||
d673344b269900a4 | zebra_multihop | multihop_counting | Consider the collection of objects shown in the image, initially viewed from the Back0 angle. You are required to perform a series of operations:
1. Remove 0 PURPLE objects.
2. Introduce 4 new RED PYRAMID objects.
3. Eliminate all RED objects present.
After these steps, determine the total count of objects remaining. | null | 6 | 1 | 4 | I must end with the total number of objects after three changes: remove no purple things, add four red pyramids, then delete every red object. The first view is from a low back angle, and the central cluster is crowded, so I cannot be certain I see every object from here. I will sketch a high, near top-down map of the ... | On this map I count seven objects: a blue cone, an orange cube, a brown pyramid, a brown cone, an orange sphere, a red cylinder and a purple cone. The first operation removes zero purple things, so nothing leaves the floor and the purple cone stays. Because the set is unchanged, I redraw the very same map, with the sam... | The map still shows the same seven objects, confirming the first step left everything in place. The second step adds four red pyramids. I will draw the same floor and the same seven objects, and place four new red pyramids in empty patches: one between the brown cone and the sphere, one to the left of the red cylinder,... | Now the map holds the original seven plus the four red pyramids I just placed. The third step erases every red object, which here means the red cylinder together with those four red pyramids. I will redraw the floor with exactly those five red pieces taken out, leaving the non-red objects where they are, so the next pi... | null | Not supported with pagination yet | The final map shows the survivors with all red pieces gone: a blue cone, an orange cube, a brown pyramid, a brown cone, an orange sphere and a purple cone. Counting them one by one gives six objects. So the total number of objects remaining is 6. | train | cc-by-nc-4.0 | procedural_grid | 21,193 | S5.replay_unsupported | |||||
a3c09d884420fe83 | zebra_multihop | multihop_counting | Given the initial arrangement of objects shown in the image below, currently viewed from the Front0 angle, execute the following sequence of operations:
Step 1: Eliminate 3 objects that are red.
Step 2: Eliminate 1 object that is red.
Step 3: Eliminate 0 objects that are spheres.
After completing these steps, determine... | null | 0 | 1 | 4 | I need the number of red cylinders left after removing three reds, then one more red, then zero spheres. From this low front view I pick out four reds: a small cone, a red cylinder, a large red cube, and a small red solid in front of the green cube; the rest I see are a green cube, a purple cylinder, and a purple spher... | The high-angle map settles the inventory: a red cone, a red cylinder, a red cube, a small red prism, a green cube, a purple cylinder, a purple sphere, and an orange torus the low view had hidden. The reds are exactly four, and the red cylinder is one of them. Since the operations remove reds four times in all, I must t... | This picture shows the scene after three reds are gone: the cone, cylinder, and cube no longer appear, and the only red left is the small red prism beside the green cube; the purple cylinder, purple sphere, and orange torus are still there. The next operation removes one red, so the prism is what goes. I will draw the ... | Now no red object appears anywhere in the picture: just the green cube, the purple cylinder, the purple sphere, and the orange torus. The last operation removes zero spheres, so nothing changes and the sphere stays. I will draw the scene once more, identical to this one, to fix the final state before counting. | null | Not supported with pagination yet | The final picture contains exactly the green cube, the purple cylinder, the purple sphere, and the orange torus. Counting reds: none of any shape remain, so in particular no red cylinder survives the removals. The total count of red cylinders remaining is 0. | train | cc-by-nc-4.0 | procedural_grid | 21,286 | S5.replay_unsupported | |||||
4fb5a19d93fbf4ab | zebra_multihop | multihop_counting | Consider the collection of objects displayed in the image below, viewed from a Right45 angle. You are instructed to perform two sequential operations:
Step 1: Remove exactly 1 BLUE object from the collection.
Step 2: Remove exactly 0 BLACK objects from the remaining collection.
After completing these operations, determ... | null | 6 | 1 | 2 | I need the number of objects left after removing one blue object and then zero black objects. In the input view I pick out seven objects: a blue sphere, a red cube, a brown cylinder in front of it, a black capsule behind it, a blue cylinder to the right, an orange capsule at the left, and a purple cube at the right. Tw... | This picture shows the scene after the removal: the blue cylinder is gone, and six objects remain where they stood - the blue sphere, the red cube, the brown cylinder, the black capsule, the orange capsule, and the purple cube. That settles the first operation. The second removes zero black objects, so nothing should c... | null | Not supported with pagination yet | null | Not supported with pagination yet | null | Not supported with pagination yet | The redrawn view is unchanged from the last one: the blue sphere, the red cube, the brown cylinder, the black capsule, the orange capsule, and the purple cube are all still present, and no other object appears. Counting them one by one, the blue sphere is one, the red cube two, the brown cylinder three, the black capsu... | train | cc-by-nc-4.0 | procedural_grid | 12,226 | S5.replay_unsupported | |||
f77874b3ca5cecc4 | zebra_multihop | multihop_counting | Given an initial scene with various objects, as depicted in the image below, we need to apply a sequence of transformations. The operations are:
1. Introduce 4 new PURPLE DONUT objects.
2. Eliminate 2 GREEN objects from the scene.
3. Remove all BLACK objects that remain.
4. Incorporate 2 BLUE DODECAHEDRON objects.
Fina... | null | 0 | 1 | 4 | I need the final count of green capsules. The input shows exactly one green capsule and one green polyhedron, among several other objects. The first operation adds four purple donuts, and later operations remove the green objects and the black ones, so I must keep the capsule distinct from the green polyhedron and plac... | This picture shows the four purple rings added: one near the green capsule, one in the middle, one near the purple pyramid, and one at the right edge. The green capsule and the green polyhedron are both still present, so these are exactly the two green objects the next operation removes. The black ring is still there, ... | In this picture the green capsule and the green polyhedron are gone, so the green removal is done and no green capsule survives it. The next operation removes all black objects. Looking across the scene, the only black object is the single ring near the top; the rest are purple, brown, or orange. So I redraw the scene ... | This picture has no black object left, so the third operation is complete. No green capsule is present, and the final operation only adds blue dodecahedra, which cannot create a green one. To reach the final scene I sketch the same viewpoint with two blue polyhedra placed near the small purple ring at the upper left, k... | null | Not supported with pagination yet | The last picture shows the two blue dodecahedra added at the top, with the unchanged set of purple, brown, and orange objects elsewhere. Scanning for any green capsule, there is none: the green capsule was removed together with the green polyhedron in the second operation, and no later step reintroduces green. Therefor... | train | cc-by-nc-4.0 | procedural_grid | 21,344 | S5.replay_unsupported | |||||
6b7ee91dbcd0c299 | zebra_multihop | multihop_counting | Given an initial set of objects shown in a top-down view, as depicted in the image, you are required to perform a series of modifications: First, eliminate all objects that are green. Second, remove all objects that are orange. Third, introduce three new blue cube objects. Finally, add three new blue cone objects. Afte... | null | 1 | 1 | 5 | The question wants the pyramid count after I delete green objects, delete orange ones, then add three blue cubes and three blue cones. From the top-down image I see a green box, a green pointed solid, an orange ball, an orange ring, a grey cylinder, and a flat dark square in the middle. That dark middle top is the trou... | This side view fixes the shapes. The dark middle solid is a pyramid, with a square base and a pointed top; one green solid is a pyramid and the other green solid is a box. There is also an orange ball, an orange ring, and a grey cylinder. I now apply the first modification, deleting everything green, which removes the ... | This picture confirms the green deletion: only the orange ball, the orange ring, the dark pyramid, and the grey cylinder are still present. The next modification deletes everything orange, so the ball and the ring must go, leaving just the dark pyramid and the grey cylinder. I will sketch that reduced scene, with nothi... | This picture shows exactly two survivors: the dark pyramid and the grey cylinder, with all green and orange objects gone. The third modification now adds three blue cubes. I will sketch the same grass with the dark pyramid and the grey cylinder still in place, plus three separate blue cubes placed around them, so that ... | In this picture I can count the three blue cubes clearly placed among the dark pyramid and the grey cylinder; the two survivors are still there. The final modification adds three blue cones. I will sketch the same scene, keeping the dark pyramid, the grey cylinder, and the three blue cubes, and placing three blue cones... | The final picture holds the dark pyramid, the grey cylinder, three blue cubes, and three blue cones. I now count pyramid shapes only. The blue pointed solids sit on round bases, so they are cones, not pyramids; the blue cubes have flat square tops; the grey cylinder is a cylinder. The only solid with a square base risi... | train | cc-by-nc-4.0 | procedural_grid | 24,337 | S5.replay_unsupported,S9.px=384 | ||||||
06de7a250499a5e3 | zebra_multihop | multihop_counting | Consider the initial arrangement of objects shown in the image below, viewed from a Back45 angle. You are required to perform two sequential operations: First, remove one DODECAHEDRON object. Second, remove all CAPSULE objects. After these operations are completed, determine the total count of red cones remaining. | null | 1 | 1 | 2 | The question asks for red cones remaining after I remove one dodecahedron, then all capsules. The input shows a red cone at lower right, a red sphere, two green dodecahedra, one central and one left, and three capsules: red, magenta, green. Two removals change the scene, so I sketch it after the first removal, from the... | In this picture the central dodecahedron is gone; what remains is the red capsule, the magenta capsule, the green capsule, the red sphere, the left green dodecahedron and the red cone. That settles the state after the first removal. The second operation removes all capsules, so I sketch the scene again from the same vi... | null | Not supported with pagination yet | null | Not supported with pagination yet | null | Not supported with pagination yet | This picture shows the scene after both removals: a red sphere at upper left, a green dodecahedron below it, and a single red cone at the right. The capsules and the extra dodecahedron are all gone, and since both operations only removed objects, no new cone appeared. Counting red cones among the three remaining object... | train | cc-by-nc-4.0 | procedural_grid | 12,178 | S5.replay_unsupported | |||
c57befa2ecd4b69c | zebra_multihop | multihop_counting | Given an initial set of 3D objects, as depicted from a Left0 angle in the image below, I need to perform a series of removal operations. First, one dodecahedron must be removed. Second, one green object needs to be removed. Third, all cube-shaped objects must be eliminated. After these steps, the final task is to deter... | null | 0 | 1 | 4 | I must report how many purple cubes remain after three removals. From this left angle I see a magenta cube at the right, a magenta dodecahedron in the middle, an orange dodecahedron beside it, red and orange cubes, a black pyramid, a green object, and a blue cylinder. The first removal says one dodecahedron, but which ... | In the overhead view I can place the whole set: two dark red cubes, an orange cube, the magenta cube at the lower right, the black pyramid, the green object, the blue cylinder, and both dodecahedrons, the magenta one sitting behind the orange one. Nothing is missing yet, so the first removal still has to be drawn. Sinc... | This picture shows the orange dodecahedron gone while the magenta one remains, so the first removal is done. The green object, the cubes, the black pyramid, and the blue cylinder are all still present, so two operations remain. The next picture should be this same overhead view with the green object removed and everyth... | Now the green object is gone as well, leaving the two dark red cubes, the orange cube, the magenta cube, the black pyramid, the magenta dodecahedron, and the blue cylinder. The only operation left is eliminating all cube-shaped objects, so I will sketch the same overhead view with exactly those four cubes removed. What... | null | Not supported with pagination yet | The last picture holds only three objects: the black pyramid, the magenta dodecahedron, and the blue cylinder. The magenta cube that stood at the lower right is gone, having been removed with the other cubes, and no other purple cube appeared in the scene. The remaining magenta solid is a dodecahedron, not a cube, so i... | train | cc-by-nc-4.0 | procedural_grid | 21,350 | S5.replay_unsupported | |||||
93050dba3c2c1267 | zebra_multihop | multihop_counting | Given the initial arrangement of objects shown in the image, viewed from the Back0 angle, perform a series of operations. First, eliminate all pyramid-shaped objects. Next, remove one object that is orange in color. Finally, introduce three new orange dodecahedron objects into the scene. After these transformations, th... | null | 1 | 1 | 4 | The question wants the number of red cones left after three edits: drop every pyramid, drop one orange object, then add three orange dodecahedrons. The input is a low back view where the small shapes crowd the horizon and sit close together, so I cannot yet be sure which solids are pyramids versus cones, or which one i... | On this top-down map I can separate the two pyramid-shaped solids: the red one at the upper right and the dark one at the lower left; the rest are cones, a cylinder, or tori, untouched yet. The first operation removes all pyramids, so I redraw the same ground with the remaining pieces in their exact spots but the two p... | With the pyramids gone, this map holds five solids: the small blue torus, the purple cylinder, the orange cone, the red cone, and the mid-right blue torus. The next operation removes one orange object. Scanning for orange, the only orange solid present is the orange cone in the middle; the cylinder is purple and the to... | This picture now shows the four survivors in place: the small blue torus up top, the purple cylinder, the red cone at the left, and the mid-right blue torus. The final operation adds three orange dodecahedrons, which are new pieces, so I keep these four exactly where they are and set three faceted orange dodecahedrons ... | null | Not supported with pagination yet | In this final scene I list the cones. The orange cone that was in the middle is no longer here, and no new cone was added, since the three new pieces are faceted dodecahedrons, not cones. The only cone remaining is the red one standing at the left; the other solids are a purple cylinder, two blue tori, and the three or... | train | cc-by-nc-4.0 | procedural_grid | 21,533 | S5.replay_unsupported | |||||
f53bcc2bda4afdb5 | zebra_multihop | multihop_counting | Consider the collection of objects shown in the image below, viewed from a Right45 angle. You are required to perform a series of operations on these objects:
Step 1: Eliminate all objects that are ORANGE in color.
Step 2: Eliminate all objects that are GREEN in color.
Step 3: Eliminate all objects that are shaped like... | null | 0 | 1 | 5 | I must track a series of removals and additions, then count what remains. In the input view I can place six objects: a green cylinder at the back left, two green cubes in the middle, a green rounded object on the right, an orange sphere at the left front, and a black cone in the center. Holding six objects through five... | In this picture the orange sphere is gone; everything else is untouched: the green cylinder, the two green cubes, the green rounded object, and the black cone, five objects in all. That settles the first removal. Next the green objects must go, and four of these five are green, so only the black cone should remain. I w... | This picture shows a single black cone on the cobblestone floor; every green object has disappeared, so the second removal is confirmed and nothing else is left. The third operation removes all cones, and the one object still present is exactly a cone. I will sketch the same viewpoint with the cone erased, which should... | The floor is now completely empty: no cylinder, cube, cone, or sphere anywhere in the view, so the first three operations are settled. The fourth operation adds two green spheres. I will sketch the same viewpoint with exactly two green spheres placed on the floor, one toward the middle and one farther to the right, so ... | This picture shows exactly two green spheres and nothing else, so the addition is confirmed: the scene now holds two objects, both spheres. The final operation removes all spheres, and every object present is a sphere, so both must disappear. I will sketch the same viewpoint with the two spheres erased, expecting a bar... | The final view shows an empty cobblestone floor with no objects anywhere in the scene. The two green spheres that were present are gone, and nothing else survived the earlier removals, so the total number of objects remaining is 0. | train | cc-by-nc-4.0 | procedural_grid | 24,256 | S5.replay_unsupported,S9.px=384 | ||||||
442761dc5b5fbe92 | zebra_multihop | multihop_counting | Consider the initial arrangement of objects presented in the image below. You are required to perform a series of removal operations in the specified order:
Step 1: Eliminate all objects identified as PYRAMIDs.
Step 2: Eliminate all objects that are BLUE in color.
Step 3: Eliminate all objects that are RED in color.
Af... | null | 1 | 1 | 3 | I need the count of orange objects after removing, in order, all pyramids, then all blue objects, then all red objects. The input shows a small orange sphere at upper left, a blue cube beside it, a small purple pyramid in the middle, a red torus at upper right, a purple sphere to its right, a red cube below the torus, ... | The picture I drew shows exactly that: the orange sphere at upper left, the blue cube beside it, the red torus and the purple sphere at the right, and the red cube in front of them. Both pyramids are gone, so the first removal is settled. What remains open is removing the blue objects, and then the red ones. So I draw ... | This picture now shows four objects: the orange sphere at upper left, the red torus and the purple sphere at the right, and the red cube in front. The blue cube has disappeared, so the second removal is done. Only the red removal is left. The red things here are the torus and the cube, so I draw the scene once more wit... | null | Not supported with pagination yet | null | Not supported with pagination yet | The last picture shows only two objects: the orange sphere at the left and the purple sphere at the right. Everything that was a pyramid, the blue cube, and both red objects are gone, so what remains is exactly these two spheres. Counting the orange ones item by item: the left sphere is orange, the right sphere is purp... | train | cc-by-nc-4.0 | procedural_grid | 16,802 | S5.replay_unsupported | ||||
09c3aa0119a1f413 | zebra_multihop | multihop_counting | Given the initial arrangement of objects shown from the Front0 angle, perform two sequential operations: first, add 4 black sphere objects, and second, remove 0 cone objects. After these operations, determine the total number of purple cylinders present. | null | 1 | 1 | 3 | I need the total number of purple cylinders after adding four black spheres and removing zero cones. The front view shows one upright purple cylinder near the center, but some objects overlap and the low angle can hide small cylinders or mislead about what is behind the cone and sphere. To make the scene unambiguous, I... | The map I sketch shows the initial scene from above, and the only purple cylinder is the upright one near the center-left; the other purple objects are a cone, a pyramid, and a faceted solid. That settles the starting count at one purple cylinder. The first operation adds four black spheres, and the second removes zero... | This picture shows the original seven objects plus four black spheres: one between the cylinder and the pyramid, one to the right of the pyramid, one near the middle left by the orange cone, and one in the lower right near the purple cone. Counting them gives exactly four added spheres. The upright purple cylinder is s... | null | Not supported with pagination yet | null | Not supported with pagination yet | The final picture is unchanged from the previous one: the four black spheres are present, and all original objects remain, including the single upright purple cylinder. The other purple objects are a cone, a pyramid, and a faceted solid, so they are not cylinders. Adding spheres and removing zero cones do not add or re... | train | cc-by-nc-4.0 | procedural_grid | 16,815 | S5.replay_unsupported | ||||
952049027cff67ba | zebra_multihop | multihop_counting | Given the initial set of objects displayed in the image, you are required to perform a sequence of four operations. First, eliminate all objects that are blue. Second, introduce four new black cylinder objects into the scene. Third, remove zero capsule objects. Fourth, remove all objects that are black. After completin... | null | 1 | 1 | 5 | I need the final number of capsules after four operations: remove all blue, add four black cylinders, remove zero capsules, then remove all black. The first view shows six objects: a red ball at the left, a blue cube behind it, an orange ball behind that, an upright orange capsule in the middle, a small blue cylinder r... | The map shows the cluster cleanly: red ball on the left, blue cube just behind it, orange ball upper right, upright orange capsule below it, and the small blue cylinder at the lower right; the blue ball from the first view is outside this frame, so it still counts. That fixes the starting set of six. Now I apply the fi... | This view holds exactly three objects: the red ball on the left, the orange ball on the right, and the upright orange capsule between and below them. All blue objects are gone, settling the first operation. The next operation adds four black cylinders. I will redraw this scene with four black cylinders placed around th... | The picture shows the three original objects plus four black cylinders: two beside the red ball on the left and two near the lower right. The addition is in place. The next operation removes zero capsules, which changes nothing, so I will redraw the identical scene, red ball, orange ball, orange capsule, and the four b... | This redraw matches the previous one exactly: red ball, orange ball, orange capsule, and the four black cylinders in the same places, so the no-op removal lost nothing. The last operation removes all black objects, meaning those four cylinders. I will draw the scene with the cylinders gone, leaving only the red ball, t... | The last picture shows only three objects: the red ball, the orange ball, and the upright orange capsule. The two round ones are spheres, not capsules; the only capsule anywhere in the scene is the upright orange one, and it survived every operation. So the number of capsules remaining is 1. | train | cc-by-nc-4.0 | procedural_grid | 24,269 | S5.replay_unsupported,S9.px=384 | ||||||
2217a72867253e73 | zebra_multihop | multihop_counting | Given an initial arrangement of six objects, as depicted in the provided image from a Left0 angle, you are tasked with performing two sequential removal operations. First, eliminate two red objects from the scene. Subsequently, remove one additional red object. After completing these steps, determine the total count of... | null | 0 | 1 | 3 | I need the number of red cylinders left after removing three red objects in two steps. The given view is low and from the left: a small rose cylinder behind a larger brown cylinder, a red cube, a purple cone partly hiding a small red lump, and a dark red pyramid. From this angle the hidden lump's shape and the true cyl... | On my map the six objects are a bright red cylinder at upper right, an orange-brown cylinder in the middle, a magenta cone at left, a small red faceted ball mid-left, a red cube at right, and a dark red pyramid at lower left. That gives four red objects and a single red cylinder; the middle cylinder is orange, not red.... | In this view the bright red cylinder and the dark red pyramid are gone; what remains is the orange-brown cylinder, the magenta cone, the red faceted ball, and the red cube. So the first operation already took the only red cylinder. The second operation removes one more red object, and I still do not know which. I will ... | null | Not supported with pagination yet | null | Not supported with pagination yet | The final view holds only three objects: the orange-brown cylinder, the magenta cone, and the small red faceted ball; the red cube has disappeared. The only cylinder still present is orange, not red, and the single red cylinder was removed in the first operation. Counting red cylinders among the remaining objects, ther... | train | cc-by-nc-4.0 | procedural_grid | 16,736 | S5.replay_unsupported | ||||
6545419ee74a51df | zebra_multihop | multihop_counting | Given an initial collection of objects displayed in the image below, perform a series of five sequential operations. After completing all operations, determine the total number of objects remaining. The operations are:
1. Eliminate all objects that are DONUTS.
2. Eliminate one SPHERE object.
3. Eliminate all objects th... | null | 2 | 1 | 5 | The question asks me to apply five operations—remove donuts, remove one sphere, remove all red, add three orange spheres, remove all spheres—and count what remains. The input shows a red torus in the upper left, an orange sphere, a small red cone, a red cylinder, a second red cone, a green cube, an orange cone, a red p... | This picture shows the scene after the donut removal: an orange sphere, two red cones, a red cylinder, a green cube, an orange cone, and a red polyhedron—seven objects, and no torus left, so operation one is done. Operation two says to eliminate one sphere, and the only sphere present is the small orange one. I will sk... | With the sphere gone, this picture holds two red cones, a red cylinder, a green cube, an orange cone, and a red polyhedron—six objects. Operation three removes every red object, which here means the cylinder, both cones, and the red polyhedron. Only the green cube and the orange cone survive, so I will sketch a view co... | This picture confirms that after the red removal only the green cube and the orange cone remain, side by side. Operation four adds three new orange spheres, so I will sketch the scene with the same cube and cone plus three orange spheres placed around them—one toward the upper left, one to the right, and one low in the... | The view now shows the green cube, the orange cone, and three orange spheres, two near the upper left and right and one low in the foreground—five objects. Operation five removes all spheres, which are exactly those three; the cube and the cone are not spheres. I will sketch the scene with the three spheres erased, lea... | The last picture shows just two objects: the green cube and the orange cone. Both survive every operation—the cube is neither a donut, a sphere, nor red, and the cone is orange, so it was never removed. With the spheres gone, nothing else remains, so the total number of objects remaining is 2. | train | cc-by-nc-4.0 | procedural_grid | 24,169 | S5.replay_unsupported,S9.px=384 | ||||||
263c713fc2aeb175 | zebra_multihop | multihop_counting | Given the initial arrangement of objects shown in the image below, viewed from the Right0 angle, you need to perform a series of operations. First, remove one DONUT object. Next, add two GREEN CONE objects. Finally, remove all objects that are PURPLE. After these steps, determine the total count of purple spheres remai... | null | 0 | 1 | 4 | The question asks how many purple spheres remain after three operations: remove one donut, add two green cones, then remove all purple objects. The input is a low, oblique view showing a purple sphere at the right, a blue pyramid, a brown cube, a green cube, and an orange donut in the middle, but the shallow angle can ... | The higher view settles the starting inventory: what looked like one donut is actually two, one tucked between the brown cube and the pyramid and one open on the right, plus the brown cube, the green cube, the blue pyramid, and the single purple sphere at the lower left. So exactly one purple sphere exists at the start... | In this picture the tucked donut is gone and only the open orange donut on the right remains, alongside the brown cube, the green cube, the blue pyramid, and the still-present purple sphere. That confirms the first operation is complete and the purple sphere has survived it. The next operation adds two green cones, so ... | This picture shows the two green cones now standing: one just in front of the brown cube and one to the left of the pyramid, with the brown cube, green cube, blue pyramid, orange donut, and the purple sphere all still present. Nothing purple has been lost yet. The final operation removes every purple object, so I will ... | null | Not supported with pagination yet | The final picture contains no purple object at all: the purple sphere that sat at the lower left has been removed, leaving only the two green cones, the brown cube, the green cube, the blue pyramid, and the orange donut. Since the last operation deletes every purple object and this picture confirms that no purple spher... | train | cc-by-nc-4.0 | procedural_grid | 21,377 | S5.replay_unsupported | |||||
95a093164ef425f2 | zebra_multihop | multihop_counting | Given an initial arrangement of various objects as seen from a Left45 angle, we need to perform two sequential operations: first, add four blue cube objects, and then add two blue cone objects. After these operations, the final question is to determine the total count of dodecahedrons present in the scene. | null | 1 | 1 | 2 | The question asks for the final number of dodecahedrons after two additions: four blue cubes, then two blue cones. In the input I see exactly one dodecahedron, the magenta one near the front left; the rest are orange capsules, black cylinders, a brown cube, and two tori. The added objects are not dodecahedrons, but I c... | In this sketch the first operation is done: I can place four blue cubes — one mid‑scene, one front left, one front center, one on the right — and the magenta dodecahedron is still there, alone and unchanged. That settles the cube step: no dodecahedron was added or covered. The second operation, two blue cones, is still... | null | Not supported with pagination yet | null | Not supported with pagination yet | null | Not supported with pagination yet | In the final view both operations are complete: two blue cones stand, one at left by the orange capsules and one at right near the brown cube, the four blue cubes remain, and the only dodecahedron is still the single magenta one. Since the added objects were cubes and cones, never dodecahedrons, and nothing occludes th... | train | cc-by-nc-4.0 | procedural_grid | 12,174 | S5.replay_unsupported,S14.sampled_qa | |||
a41563217ec0e5a9 | zebra_multihop | multihop_counting | Given the initial set of objects displayed in the image below, viewed from the Back0 perspective, perform the following two sequential operations:
Step 1: Introduce 2 additional BLUE PYRAMID objects to the scene.
Step 2: Introduce 4 additional BLACK CONE objects to the scene.
After completing these steps, what is the t... | null | 13 | 1 | 3 | The question asks for the total number of objects after adding two blue pyramids and then four black cones. In the input view I can identify seven objects: a brown metal cone, a matte green polyhedron, two small purple toruses, a green pyramid, a black cone, and a large purple cone. But this low-angle view will get cro... | The map shows exactly the seven original objects: the green polyhedron and green pyramid at the top, the two purple toruses between them, the brown cone at the left, and the black cone and purple cone at the right. That fixes the base count at seven. What is still open is step 1: placing the two new blue pyramids. I wi... | This picture now shows nine objects: the seven originals plus the two blue pyramids, one at the top center and one at the bottom center. That settles step 1. Step 2 remains: adding four black cones. I will redraw the same view with four new black cones placed in the open spaces around the middle of the scene, keeping a... | null | Not supported with pagination yet | null | Not supported with pagination yet | In the final view I count item by item: the green polyhedron, the green pyramid, the two purple toruses, the brown cone, and the purple cone, six objects; the two blue pyramids; and the black cones, one original plus four new, making five. So six plus two plus five gives 13. The scene therefore contains 13 objects, and... | train | cc-by-nc-4.0 | procedural_grid | 16,774 | S5.replay_unsupported | ||||
5b3931f3cdd88faf | zebra_multihop | multihop_counting | Consider the arrangement of objects shown in the image below, viewed from a Left45 angle. You are required to perform two sequential operations on these objects:
Step 1: Eliminate one ORANGE object from the scene.
Step 2: Introduce three new BLUE CAPSULE objects into the scene.
After completing these steps, determine t... | null | 2 | 1 | 2 | I need the final number of green objects after removing one orange object and adding three blue capsules. The input view shows an orange cylinder at the right, a green cone upper right, a green cylinder lower left, and blue cones elsewhere. The operations are not yet visible, and I must make sure the orange object is r... | In this picture the orange cylinder is gone. The green objects visible are the green cone at upper right and the green cylinder at lower left; the other remaining objects are blue cones. The second operation has not been shown yet: three blue capsules must be introduced. Since capsules are blue they cannot add green ob... | null | Not supported with pagination yet | null | Not supported with pagination yet | null | Not supported with pagination yet | The final picture shows the scene after both operations: the green cone at upper right, the green cylinder at lower left, the blue cones, and three blue capsules in the middle. I count green objects one by one. The green cone is green. The green cylinder is green. The blue cones and the blue capsules are not green. No ... | train | cc-by-nc-4.0 | procedural_grid | 12,242 | S5.replay_unsupported | |||
ce96ce691a027af2 | zebra_multihop | multihop_counting | Given an initial arrangement of objects as seen from the Front0 angle in the image below, perform a series of modifications:
1. Eliminate all black-colored objects.
2. Eliminate all orange-colored objects.
3. Introduce two green cone-shaped objects.
After these operations, how many red objects remain in the scene? | null | 2 | 1 | 4 | The question asks how many red objects remain after removing black objects, removing orange objects, and adding two green cones. The front view shows the red cone, the black cube and cylinder, the orange triangular prism, the green torus, the brown torus and capsule, and a blue shape behind the black cube. But this low... | In this high view I can see the whole scene. Where the black cube hid it from the front, there is a red dodecahedron, and a blue cube stands in front of the black cylinder. So the red objects present are the red cone on the left and the red dodecahedron in the middle. The black objects are the black cube and the black ... | The view now has the black cube and black cylinder gone, so the first operation is applied. What remains: the red cone, the red dodecahedron, the green torus, the blue cube, the brown torus, the brown capsule, and the orange triangular prism at the right. The second operation removes all orange objects, so the orange p... | In this picture the orange prism is gone, so the second operation is applied; the brown torus and brown capsule are not present either, so I take this picture as the current scene. What remains is the green torus, the red cone, the red dodecahedron, and the blue cube. No red object was touched by the removals. The thir... | null | Not supported with pagination yet | The final view shows the completed scene: two new green cones at the top, plus the green torus, the red cone, the red dodecahedron, and the blue cube. Counting red objects one by one: the red cone on the left and the red dodecahedron in the middle. The cones that were added are green, and the removals only took out bla... | train | cc-by-nc-4.0 | procedural_grid | 21,409 | S5.replay_unsupported | |||||
356c374e4b539f5c | zebra_multihop | multihop_counting | Given the initial arrangement of objects shown in the image, perform a series of transformation steps. First, add two orange dodecahedron objects. Next, remove all cube-shaped objects. Then, introduce two black sphere objects. Following that, add two purple sphere objects. Finally, remove zero purple objects. After com... | null | 1 | 1 | 5 | I need the final count of red objects after a sequence of edits. The input shows one red object: a small red pyramid sitting in the center. The edits add orange dodecahedrons, delete cubes, add black spheres, add purple spheres, and remove zero purple pieces; none of these mention red, but I will not trust memory, so I... | This picture shows the two orange dodecahedrons placed, and the red pyramid still sitting in the center, untouched by the addition. That settles the first edit. The next edit removes every cube-shaped object, so the blue cube on the left and the purple cube on the upper right should vanish. I redraw the scene without t... | In this picture the two cubes are gone and no red piece was lost; the red pyramid remains the only red object in the center. That settles the cube removal. The next edit introduces two black spheres. I sketch the same layout and add two black spheres, one to the left of the small top dodecahedron and one to the right o... | Here the two black spheres sit where I put them, and the red pyramid is unchanged, still the sole red object. That settles the sphere addition. The following edit adds two purple spheres. I draw the same scene plus two purple spheres, one in the middle above the central dodecahedron and one on the right above the brown... | This picture shows the two purple spheres in place and the red pyramid still present; the purple step created no red object. That settles the fourth edit. The final edit removes zero purple objects, so the scene should stay exactly as it is now. I redraw it unchanged, the same objects in the same positions, so I can re... | In the final picture I scan for red: the spheres are black and purple, the added dodecahedrons orange, the left pyramid blue, the cylinders green and purple, and the tori green and brown; the only red thing left is the single red pyramid in the center. Since no edit ever added or removed red, that starting red pyramid ... | train | cc-by-nc-4.0 | procedural_grid | 24,263 | S5.replay_unsupported,S9.px=384 | ||||||
e12ef84a1ea5d446 | zebra_multihop | multihop_counting | Given an initial arrangement of objects as seen in the image, you are instructed to perform two sequential operations: First, remove two cone-shaped objects. Second, add two new purple cylinder-shaped objects to the scene. After these operations, determine the total count of purple-colored objects present. | null | 2 | 1 | 2 | The question asks for the total number of purple objects after two operations: first remove two cone-shaped objects, then add two purple cylinders. The input image shows six objects: a red capsule, a red pyramid, a red cube, a blue cone, a green cone, and a brown cone. The input alone does not show the scene after the ... | The view I sketched shows the first operation done: the green cone and the brown cone are gone, and the red capsule, red pyramid, red cube, and blue cone remain in their original places. That settles the removal step. What is still open is the second operation, adding two purple cylinders. No purple object exists in th... | null | Not supported with pagination yet | null | Not supported with pagination yet | null | Not supported with pagination yet | This picture shows the scene after both operations. Counting purple objects: one purple cylinder with a magenta top stands to the left of the red cube, and a second purple cylinder stands in front of it; the red capsule, red pyramid, red cube, and blue cone are not purple. So the only purple objects are the two cylinde... | train | cc-by-nc-4.0 | procedural_grid | 12,203 | S5.replay_unsupported,S8.phash_near_but_distinct | |||
68bbf7ba45ab0c9b | zebra_multihop | multihop_counting | Given an initial arrangement of objects shown in the image below, perform a series of five sequential modifications. After completing all operations, determine the total count of green objects remaining.
The operations are:
1. Introduce 2 BLACK CAPSULE shapes.
2. Eliminate 1 BLACK object.
3. Incorporate 4 RED SPHERE sh... | null | 2 | 1 | 5 | I need the number of green objects left after five operations. In the input I see four green objects: a green cube, a small green cone beside the red cylinder, a green cone at the left, and a green sphere at the bottom. The operations add black capsules, remove a black object, add red spheres, discard three cones and r... | The map I sketch shows the two new dark capsules: one standing in the upper left and one in the middle between the green sphere and the magenta ball. The two black cones still flank the red cylinder, and none of the green objects changed. Still open is the second operation, removing one black object, and the words do n... | This picture shows the left black cone gone while the one at the right of the red cylinder remains, and both dark capsules stay put. So the removal took a black cone. The cones now present are the small green cone, the left green cone, and that black cone. The next operation adds four red spheres, and their positions a... | The drawn view now has four new red spheres: one above the green cube, one further to the left, and two to the right of the red cylinder, in addition to the original small red one. The addition is settled. Next, three cones are discarded, and the only cones visible are the small green cone, the left green cone, and the... | This picture has no cones left, so the green cube and green sphere are the only green objects still standing. The last operation removes one cylinder, and the only cylinder in view is the red one, so it is the piece that goes. I will draw the same view without the red cylinder to get the final arrangement. | The final picture shows the red cylinder gone and, in green, only the green cube and the green sphere. Counting them one by one: the cube, then the sphere, that is two green objects; both green cones left with the cone removal and nothing green was ever added. The answer is 2. | train | cc-by-nc-4.0 | procedural_grid | 24,265 | S5.replay_unsupported,S9.px=384 | ||||||
bedce6d8d773315e | zebra_multihop | multihop_counting | Given the initial set of objects shown in the image below, observed from a Front45 perspective, I need to perform two sequential operations: first, add two blue dodecahedron shapes, and second, add two black cone shapes. After completing these additions, the task is to determine the total count of black objects present... | null | 3 | 1 | 2 | The question asks for the black object count after adding two blue dodecahedrons, then two black cones. The input shows one black object, a cube at upper middle, plus a red cylinder, magenta cylinder, orange sphere, magenta cone, and magenta pyramid. The additions are described but not visible, so the state after each ... | The sketched view shows the first addition done: the six original objects remain, red cylinder, black cube, magenta cylinder, orange sphere, magenta cone, magenta pyramid, and two blue dodecahedrons now sit in the scene, a small one near the middle and a larger one at lower right. Thus the black count is still one, the... | null | Not supported with pagination yet | null | Not supported with pagination yet | null | Not supported with pagination yet | This picture shows the final state: next to the magenta cone stand two black cones, one just behind the other, and the black cube remains at upper middle. Counting black objects one by one, the cube, the nearer black cone, and the farther black cone, that is three. The two blue dodecahedrons, the red cylinder, the mage... | train | cc-by-nc-4.0 | procedural_grid | 12,259 | S5.replay_unsupported | |||
622702586d50b050 | zebra_multihop | multihop_counting | Given an initial set of objects displayed in the image below, viewed from the Back0 perspective, perform two sequential operations. First, add two orange cube objects to the scene. Second, remove all green objects present. After these operations, determine the final count of red spheres. | null | 1 | 1 | 3 | The question asks for the number of red spheres after adding two orange cubes and removing all green objects. The input shows a low-angle view with a red sphere left, a red cylinder and small red cone, a large orange capsule center, a small green cone, a magenta polyhedron, a purple cone, and a green ring right. The ca... | In this top-down view every object stands apart on the ground: the red sphere at bottom center, the red cylinder and small red cone at bottom left, the orange capsule with the green ring beside it on the right, the small green cone just below the magenta polyhedron, and the purple cone up top. Nothing is hidden behind ... | This picture shows the two added orange cubes in place: one right beside the purple cone and one in the open ground at the center, with the rest of the starting set unchanged, including the green ring by the capsule and the small green cone below the magenta polyhedron. The addition is settled; what remains is the remo... | null | Not supported with pagination yet | null | Not supported with pagination yet | In this final picture the green ring and the small green cone are gone: the capsule now sits alone on the right and the magenta polyhedron stands without the cone below it, while the two orange cubes, the purple cone, the red cylinder, the red cone, and the red sphere remain. Counting red spheres among these: the red c... | train | cc-by-nc-4.0 | procedural_grid | 16,857 | S5.replay_unsupported | ||||
a0b49829b60ade49 | zebra_multihop | multihop_counting | Given an initial set of objects presented in the image below, viewed from a Back45 angle, perform a series of operations. First, eliminate all objects classified as CUBEs. Second, remove zero SPHERE objects. Finally, determine the total count of black cubes remaining after these operations. | null | 0 | 1 | 2 | I must count black cubes after two operations: delete every cube, then delete zero spheres. In the input I can see one black cube sitting to the right of the green ring, and one blue cube in the center; both are cubes and both would be deleted by the first operation. Holding that in my head while also tracking the sphe... | The map I sketched shows the scene with both cubes gone: I see the brown capsule, the brown sphere, the green ring, the green sphere, the blue ring, and the red polyhedron, and no cube of any colour remains among them. That settles the first operation. The second operation removes zero spheres, so nothing else is taken... | null | Not supported with pagination yet | null | Not supported with pagination yet | null | Not supported with pagination yet | This final picture contains no cube at all; the only solid blocks that existed were the blue one and the black one, and both have already been erased. With the cube set empty, there is nothing black and cubed left to count. Therefore the total number of black cubes remaining after the operations is 0. | train | cc-by-nc-4.0 | procedural_grid | 12,125 | S5.replay_unsupported | |||
2c79e972f5ece3df | zebra_multihop | multihop_counting | Consider the initial arrangement of objects shown in the image, viewed from the Back0 angle. Your task is to perform two sequential operations: First, eliminate all blue objects. Second, remove zero green objects. Finally, determine the total count of blue cylinders remaining after these operations. | null | 0 | 1 | 3 | The question asks me to eliminate all blue objects, then remove zero green objects, and count the blue cylinders left. The input view is the Back0 angle it names: on the left a black rounded object partly behind a purple capsule, in the middle a green capsule, on the right a blue cylinder with a red pyramid and a small... | The map I sketch shows six objects: the blue cylinder, the green capsule, the purple capsule, the black rounded object, the red pyramid, and the orange torus. The only blue thing is that single blue cylinder, so the first operation removes exactly it. The second operation removes zero green objects, so the green capsul... | This picture shows the scene after the blue objects are gone: the green capsule, the purple capsule, the black rounded object, the red pyramid, and the orange torus remain, and the spot where the blue cylinder stood is now empty. Since the second operation removes zero green objects, nothing else changes. I sketch the ... | null | Not supported with pagination yet | null | Not supported with pagination yet | In the final picture, the green capsule, the purple capsule, the black rounded object, the red pyramid, and the orange torus are present, and no blue object appears anywhere. The only blue cylinder that existed was removed in the first operation, and the second operation changed nothing. So the number of blue cylinders... | train | cc-by-nc-4.0 | procedural_grid | 16,730 | S5.replay_unsupported | ||||
cf37bb405897828f | zebra_multihop | multihop_counting | Given an initial setup of various 3D objects as depicted in the image below, you are required to perform a sequence of four operations. First, remove all pyramid-shaped objects. Second, add four new red capsule objects to the scene. Third, remove all capsule-shaped objects that are present. Fourth and finally, introduc... | null | 2 | 1 | 4 | I must count the donut-shaped objects left after four operations: remove pyramids, add four red capsules, remove all capsules, add two blue dodecahedrons. The input shows a blue cylinder, a red pyramid at the right, a blue capsule at the right center, a purple donut at the left, a red donut in the middle, and a blue py... | The map I sketch shows exactly four objects now: the blue cylinder where it stood, the blue capsule to its right, and the purple and red donuts in the middle. Both pyramids are gone, so the first operation is settled. Next I add four red capsules. I redraw the same viewpoint with the four objects unchanged and four shi... | In this picture I count the red capsules one by one: upper left, upper right, lower left of the red donut, and lower right — four of them, matching what I added, plus the original blue capsule. The third operation removes every capsule, so all five of these go: the blue one and the four red ones. What stays is the blue... | This picture shows the scene after the capsule removal: the blue cylinder at the top, the purple donut and red donut beneath it, and nothing else. Every capsule is gone, as the operation required. The last operation adds two blue dodecahedrons. I sketch the same view with the cylinder and both donuts unchanged, and two... | null | Not supported with pagination yet | The final picture shows the blue cylinder, two blue dodecahedrons at the right and bottom center, and exactly two donuts: the purple one at the left and the red one in the middle. Counting them item by item: one purple donut, one red donut. No operation ever added or removed a donut, so these are all that remain. That ... | train | cc-by-nc-4.0 | procedural_grid | 21,424 | S5.replay_unsupported,S8.phash_near_but_distinct | |||||
3a6cbaeca3e9bbd4 | zebra_multihop | multihop_counting | Given an initial setup of various 3D objects, as depicted in the image below, which is viewed from the Right0 angle, you need to perform a sequence of operations:
Step 1: Eliminate 0 PURPLE objects.
Step 2: Eliminate 2 CUBE objects.
After these operations are completed, determine the total count of pyramid-shaped objec... | null | 2 | 1 | 3 | The question asks me to first remove 0 purple objects, then 2 cubes, and finally count the pyramids that remain. From the given right-side view I can see a purple cylinder, a small red pyramid, a brown cube, a red torus, a purple cube, and a dark pyramid tucked behind the purple cube; the occlusion makes it hard to be ... | On my sketch I can now read off all six objects without occlusion: the purple cylinder, the small red pyramid, the brown cube, the red torus, the purple cube, and the black pyramid standing clear of the purple cube. That settles the inventory: the cubes are exactly the brown and the purple one, and the pyramid shapes a... | The redrawn scene is unchanged from my sketch: the purple cylinder, red pyramid, brown cube, red torus, purple cube, and black pyramid are all still there, so the first operation indeed removed nothing. The second operation removes the two cubes, which in this picture are the brown cube at the back and the purple cube ... | null | Not supported with pagination yet | null | Not supported with pagination yet | In this final picture only four objects remain: the purple cylinder, the red torus, the red pyramid, and the black pyramid. The two cubes are gone, and the cylinder and torus are not pyramid-shaped. Counting the pyramid shapes item by item, there is the red pyramid and the black pyramid, which gives a total of 2. So th... | train | cc-by-nc-4.0 | procedural_grid | 16,827 | S5.replay_unsupported | ||||
6bff82eda5ddf9be | zebra_multihop | multihop_counting | Given the initial arrangement of objects shown in the image below, perform a series of operations:
Step 1: Incorporate 3 PURPLE CONE objects into the scene.
Step 2: Eliminate 3 objects that are PURPLE in color.
Step 3: Introduce 4 BLACK DONUT objects.
Step 4: Add 3 BLUE DODECAHEDRON objects.
After completing all these ... | null | 0 | 1 | 4 | I need the final number of purple donuts. The first image shows one purple donut, a red donut, two orange dodecahedra, a brown cylinder, and a red cylinder, with no purple cones. The operations add purple cones and then remove three purple objects, so the input alone does not say whether the purple donut survives that ... | This picture shows the original purple donut plus three purple cones: the cone above it, the cone by the upper orange shape, and the tilted cone at lower left, making four purple objects. The next operation removes three purple objects, but this view does not yet say which three disappear. Since the question depends on... | This view now lacks the purple ring from the first image; one purple cone stands near the middle. The red donut, two orange dodecahedra, brown cylinder, and red cylinder remain. That shows the removal step took the purple donut. The later additions are black donuts and blue dodecahedra, which are not purple rings, so I... | This picture shows four black donuts: one above the red donut, one beside the purple cone, one lower left, and one partly behind the red cylinder. It still lacks that purple ring; only the single purple cone remains purple. The final operation adds three blue dodecahedra, which are not purple rings, so I need to confir... | null | Not supported with pagination yet | The completed view contains three blue dodecahedra at upper right, middle right, and lower right, along with four black donuts, a red donut, two orange dodecahedra, a purple cone, a brown cylinder, and a red cylinder. Counting purple donuts specifically, there is no purple ring or torus anywhere in this final scene; th... | train | cc-by-nc-4.0 | procedural_grid | 21,441 | S5.replay_unsupported,S8.phash_near_but_distinct | |||||
ad87f727b5d9e20a | zebra_multihop | multihop_counting | Given the initial arrangement of objects shown in the image, viewed from a Right45 angle, perform a series of operations. First, add two black cube objects. Next, remove one orange object. Then, remove two black objects. Finally, remove one donut object. After completing all these steps, determine the total count of bl... | null | 1 | 1 | 4 | I need the final count of blue cubes after four operations: add two black cubes, remove one orange object, remove two black objects, remove one donut. In the input I see one blue cube, plus a black sphere, a green capsule, a green cylinder, an orange cone, and a black torus. The input alone does not settle it, since th... | This picture shows the first operation done: all six original objects remain, and two black cubes have joined, one beside the green cylinder at the upper right and one on the open floor at the lower left. The black objects are now the sphere, the torus, and the two cubes. The orange removal is still pending, so I redra... | The orange cone is gone, so the orange removal is settled. What remains: the black sphere, the green capsule, the green cylinder, the blue cube, the black torus, and the two black cubes. Next I must remove two black objects; the sphere is the most isolated, so I remove it first. I redraw the same view without the spher... | With the sphere gone, the black objects left are the torus and the two black cubes. I still owe one more black-object removal and the donut removal. The donut is the torus, so it disappears next; for the second black object I take the black cube at the upper right, keeping the one at the lower left. I redraw the view w... | null | Not supported with pagination yet | The final picture shows exactly four objects: the green capsule, the green cylinder, the blue cube, and one black cube at the lower left. Counting blue cubes in this view, there is only the single cube in the center; none of the operations added or removed blue cubes, and the rest are green or black. So the total count... | train | cc-by-nc-4.0 | procedural_grid | 21,330 | S5.replay_unsupported | |||||
43d6019306eea829 | zebra_multihop | multihop_counting | Given an initial arrangement of objects viewed from a top-down perspective, perform a series of four operations:
1. Remove 0 GREEN objects.
2. Remove 0 GREEN objects.
3. Add 4 ORANGE DODECAHEDRON objects.
4. Remove 1 ORANGE object.
After completing these steps, the task is to determine the total number of dodecahedrons... | null | 4 | 1 | 5 | The question asks how many dodecahedrons remain after four operations. From directly above I see a purple cylinder, a green cylinder, an orange torus, and two orange round shapes, one middle and one lower left. The flat view cannot tell whether those two shapes are cones or spheres, and the operations add orange dodeca... | This oblique view settles the shapes: the middle orange and the lower-left orange objects are both cones, with pointed tops, and the left orange object is the torus; the cylinders are purple and green. No dodecahedron is present at the start. The first operation removes 0 green objects, so nothing in the scene changes.... | The redrawn scene is identical to the last: the orange torus at left, the two orange cones, and the purple and green cylinders, and still no dodecahedron. That confirms the first removal changed nothing. The second operation also removes 0 green objects, so the scene stays the same. I will sketch this same view once mo... | In this picture the scene is still just the torus, the two orange cones, and the two cylinders; there is no dodecahedron yet. The third operation adds 4 orange dodecahedrons, so the next drawing must contain the same five objects plus four faceted orange dodecahedrons on the empty grass: one above the middle cone, one ... | This picture shows the four added dodecahedrons: counting them, one at the top, one at the left, one to the right of the green cylinder, one at the lower right. Alongside them remain the orange torus, the two orange cones, and the two cylinders. The last operation removes 1 orange object. I will remove the orange torus... | In this final picture the orange torus is gone, while the two orange cones, the purple cylinder, the green cylinder, and the faceted orange dodecahedrons remain. Counting the dodecahedrons item by item: the one at the top, the one at the left, the one to the right of the green cylinder, and the one at the lower right. ... | train | cc-by-nc-4.0 | procedural_grid | 24,297 | S5.replay_unsupported,S9.px=384 | ||||||
7498f384f70bb8ba | zebra_multihop | multihop_counting | Consider the collection of objects shown in the image below, initially viewed from the Right0 perspective.
Perform the following two sequential operations:
Step 1: Remove two objects that are black in color.
Step 2: Add two new objects, both of which are black cylinders.
After these operations are completed, determine ... | null | 2 | 1 | 3 | I must count the purple objects after two operations: removing two black objects, then adding two black cylinders. From this right view I see a purple dodecahedron at the front left, a purple cylinder in the back, a brown cylinder, a red dodecahedron, and a black cylinder at the front right. The black cylinder's body m... | The map shows six objects on the floor: the purple cylinder and the brown cylinder in the back, a dark sphere to the upper right of the red dodecahedron, the red dodecahedron, the purple dodecahedron, and the black cylinder at the front right. So there are exactly two black objects, the black cylinder and the dark sphe... | This picture confirms the removal: only the purple cylinder, the brown cylinder, the red dodecahedron, and the purple dodecahedron remain; both black objects are gone. The second operation adds two black cylinders, so I draw the same scene with two black cylinders inserted, one on the open floor to the left and one bet... | null | Not supported with pagination yet | null | Not supported with pagination yet | The final picture shows six objects: two black cylinders, the brown cylinder, the red dodecahedron, the purple dodecahedron, and the purple cylinder. The removals touched only black objects and the additions bring only black ones, so the purple objects are unchanged from the start. Counting the purple ones: the purple ... | train | cc-by-nc-4.0 | procedural_grid | 16,777 | S5.replay_unsupported | ||||
76acc00993c38dfa | zebra_multihop | multihop_counting | Given the initial arrangement of objects shown in the image below, perform a series of operations:
1. Add 3 ORANGE PYRAMID objects.
2. Remove 1 ORANGE object.
3. Remove 2 ORANGE objects.
4. Remove all BLUE objects.
After completing these steps, determine the total count of red objects remaining. | null | 1 | 1 | 4 | The question asks how many red objects remain after four operations. In the input I see one red object, a small red torus on the right, one orange capsule, four blue objects (a capsule, two cylinders, a cube) and a green square. The operations add and remove only orange and blue pieces, so the torus should survive, but... | The sketch shows the full starting set: the orange capsule plus three orange pyramids, so four orange objects, with the red torus untouched. The next operation removes one orange object. I will redraw the scene with the orange pieces reduced, keeping the orange capsule standing and the pyramid in front of the blue cube... | In this picture the orange capsule still stands and one orange pyramid sits in front of the blue cube; the side pyramids are gone, which is one more removal than the last operation alone, so I take the picture as the true state. The blue set is unchanged: capsule, two cylinders, cube, and the red torus is still there. ... | This picture shows exactly one orange object left, the pyramid in front of the blue cube, with the capsule gone, while all four blue objects are still present. The last operation removes every blue object. I will redraw the scene from the same viewpoint with the blue capsule, both blue cylinders and the blue cube erase... | null | Not supported with pagination yet | The final picture shows three objects: the orange pyramid, the green square, and the red torus. No blue object remains, as the last operation required. Scanning for red, only the torus is red, and none of the operations added or removed it. So the total count of red objects remaining is 1. | train | cc-by-nc-4.0 | procedural_grid | 21,334 | S5.replay_unsupported,S8.phash_near_but_distinct,S14.sampled_qa | |||||
9bdbd5a7ed7249a7 | zebra_multihop | multihop_counting | Given an initial arrangement of objects as shown in the image below, viewed from the Left0 angle, you are asked to perform a sequence of operations.
The operations are:
Step 1: Remove 0 PURPLE objects.
Step 2: Remove all ORANGE objects.
After performing these steps, determine the total count of black objects remaining. | null | 3 | 1 | 3 | The question asks for the number of black objects left after removing zero purple objects and then all orange objects. In the input view I can place a black cylinder, a black sphere, and a small black flat torus, alongside a red cube, a green torus, and a purple cube, with an orange tip poking up behind the cylinder. T... | This higher view settles it: the orange piece is one cone standing just behind the black cylinder, and nothing else is hidden. I can now list every object: green torus front left, red cube in the middle, black cylinder, orange cone behind it, purple cube on the right, black sphere in front of the cube, and a small blac... | The redrawn view matches the previous one exactly, as expected after a zero-removal step; the purple cube and all other pieces are where they were. The open question is now step 2, which deletes every orange object. There is only one: the cone behind the cylinder. Everything else must remain. So I draw the same high vi... | null | Not supported with pagination yet | null | Not supported with pagination yet | With the cone gone, this last view is the scene after both operations. The dark objects on the floor are the black cylinder, the black sphere, and the small black flat torus. Counting them one by one: the cylinder is one, the sphere is two, the flat torus is three. The purple cube, red cube, and green torus are not bla... | train | cc-by-nc-4.0 | procedural_grid | 16,797 | S5.replay_unsupported | ||||
a0b5ea8bb3c99288 | zebra_multihop | multihop_counting | Given the initial arrangement of objects shown in the image, currently viewed from the Back0 angle, perform a sequence of two operations: First, add three GREEN DONUT objects to the scene. Second, remove all objects that are GREEN. After these operations are completed, determine the total number of blue pyramids remain... | null | 1 | 1 | 3 | I must count the blue pyramids left after two changes: first adding three green donuts, then deleting every green object. The ground-level view shows one blue pyramid, a black cylinder, a brown and a purple donut, and red, green and orange cylinders, but at this angle it is hard to keep track of where new donuts land a... | The map I sketch puts the blue pyramid near the middle, with the black cylinder, brown donut and purple donut to its lower left and the red, green and orange cylinders to its upper left. Nothing has changed yet. The first operation adds three green donuts, so I redraw this same top-down map and place three green rings ... | This picture now shows the three new green rings on the right in addition to the original green cylinder at the top, so the green objects are that cylinder plus the three rings. The second operation removes everything green, so I redraw the map with the green cylinder and the three green rings erased, keeping the blue ... | null | Not supported with pagination yet | null | Not supported with pagination yet | In this final map the green cylinder and all three green rings are gone; what remains is the blue pyramid, the black cylinder, the red cylinder, the orange cylinder, the brown donut and the purple donut. The blue pyramid was never green and no operation touched it, so exactly one blue pyramid is still present. The answ... | train | cc-by-nc-4.0 | procedural_grid | 16,735 | S5.replay_unsupported | ||||
79d305e76c4b2f85 | zebra_multihop | multihop_counting | Given the initial arrangement of objects shown in the image below, which is viewed from the Back0 angle, apply the following sequence of transformations: First, introduce 2 RED SPHERE objects. Next, add 2 GREEN CUBE objects to the scene. Subsequently, eliminate all objects that are BLUE in color. Finally, remove a sing... | null | 2 | 1 | 5 | I need the number of purple objects left after adding two red spheres, adding two green cubes, removing every blue object, and removing one black object. In the first view I see two purple objects: a faceted polyhedron and a small pointed one. But the tall black cylinder hides what stands right behind it, with only a s... | The map I sketch shows the layout from above: a black capsule at the left, the purple polyhedron beside the black cylinder, a small blue ring and a blue cylinder to its right, a blue polyhedron in front, the orange cone at the lower right, and two objects the first view hid, a red polyhedron and a wedge-shaped purple f... | This picture now includes the two red spheres: one sits at the upper right, past the orange cone, and the other lies to the left of the blue polyhedron. The purple pair is unchanged, and the red additions do not touch the blue or black objects that come next. The next operation adds two green cubes, so I will redraw th... | In this picture the two green cubes sit side by side in the foreground, one to the left of the other. Everything else holds its place: the black capsule, the purple polyhedron, the black cylinder, the blue ring, the blue cylinder, the blue polyhedron, the two red spheres, the red polyhedron, the purple wedge, the orang... | The blue objects are gone: the ring, the cylinder, and the polyhedron no longer appear. What remains are the black capsule, the black cylinder, the purple polyhedron, the purple wedge, the red polyhedron, the two red spheres, the orange cone, and the two green cubes. The final operation removes a single black object, s... | In this last picture the black cylinder is gone while the black capsule remains, so the single black removal is complete. Counting the purple objects now visible: the faceted purple polyhedron near the upper middle, and the wedge-shaped purple object on the right. That is two purple objects, and none of the operations ... | train | cc-by-nc-4.0 | procedural_grid | 24,336 | S5.replay_unsupported,S9.px=384 | ||||||
e8c6e12b1006620d | zebra_multihop | multihop_counting | Given an initial setup of objects as depicted in the image below, you are asked to perform a series of sequential operations. The initial view is from a Right45 angle.
The operations are:
Step 1: Introduce 2 additional BLUE CYLINDER objects to the scene.
Step 2: Eliminate all objects that are BLUE in color.
Step 3: Inc... | null | 0 | 1 | 3 | The question asks for the number of blue cylinders left after three operations: add two blue cylinders, remove every blue object, add four green capsules. In the input view I see one blue cylinder in the front, plus a brown cylinder, a black prism, a purple sphere, and a magenta pyramid. One view can't carry the whole ... | In the map I sketch, three blue cylinders now stand in the scene: the original light-blue one in the front, the one left of the magenta pyramid, and the one beside the brown cylinder. The other objects stay brown, black, purple, and magenta, so no other object is blue. That settles the state after the first operation. ... | This picture shows the result of the removal: the brown cylinder, the black prism, the purple sphere, and the magenta pyramid remain on the ground, and no blue object appears anywhere in the view. So the removal has taken every blue cylinder away. The third operation adds green capsules, which are not blue, but I want ... | null | Not supported with pagination yet | null | Not supported with pagination yet | In the final picture I look for blue objects one by one: there is the brown cylinder, the black prism, the purple sphere, the magenta pyramid, and four green capsules, and no blue cylinder stands anywhere in the scene. The original blue cylinder and the two added ones were all blue, so the removal step took every one o... | train | cc-by-nc-4.0 | procedural_grid | 16,899 | S5.replay_unsupported |
End of preview. Expand in Data Studio
Spatial MMCoT v1 · zebra_multihop
Zebra-CoT multi-hop object counting over rendered 3D scenes (primitive objects on textured ground under a sky). Many traces open with a viewpoint change, so the first target is often a novel view. The final thought re-examines the last target image, but on some rows it only says it counts the objects and never states the number; the count is then only in <answer>. Released rows carry 2 to 5 target images. 2,057 traces with more target images than the declared maximum were refused (S9), not cut, so every released row keeps its full chain. 707 rows (flags S9.px=*) were over the token budget at full size, so their targets are stored smaller; the BAGEL/ThinkMorph VAE transform (min_image_size 512) scales them back up for the image loss, so those targets are trained as upsampled, softer images. The thoughts keep upstream's numbered labels ('THOUGHT 0:', 'THOUGHT 1:', ...) inside <think>. They count upstream's thoughts across the whole trace, not the row's slots, so one slot often holds two or more labelled thoughts, and a slot after a target image can open with the unlabelled end of the previous thought. No other source in this release has these labels; if you mix sources and do not want a model to learn them, remove them with re.sub(r'THOUGHT \d+:\s*', '', text). Refused by the conversion's checks: 1,142 rows whose question announces operations without listing them (S0.ops_not_stated); 2,061 rows whose answer is not an integer (S0.non_integer_answer); 33 rows whose read-back is empty (S0.empty_readback).
Reasoning text in this release (r4)
The paragraph above describes the upstream data, and where it describes the reasoning text it describes the text of the previous release (r3). In this release every thought was rewritten: Qwen3.8-27B in thinking mode (temperature 1.0, top-p 0.95, top-k 20; prompt v4.1, scripts/annotate/rationale.py in our conversion code (the code repository is not public yet)) was shown the row's input images, its target images in order, the question, the options and the answer, and wrote the thought before each target image (what is still unknown, and what the picture should show to settle it) and the read-back after the last one. The images, the question, the options, the answer and the number of steps are r3's, checked row by row (image bytes, instruction, K and the <answer> value).
This is the training subset: 112 rows of r4 are not in it, listed with their reason in reports/subset_dropped.jsonl: 77 the rewritten row's estimated token cost exceeds the 24,576-token sample budget, which the trainer's packer would drop silently (over_budget); 35 the rewrite model judged the stored answer doubtful (its note is in the list) (label_doubt).
Self-check of the rewritten text (measured 2026-10-02 on 22 rows sampled from this source before the removals above, each judged against its full-resolution pictures): 0.95 faithfulness errors per row (a claim the pictures contradict), 55% of rows with none; the plan names the real obstacle on 77%, the read-back reads the picture it follows on 86%, and the answer follows from the text on 96%.
The measured caveats on the r3 card were counted on r3's rows and text and are not repeated here; the per-row known-issue lists below were measured again on these rows and this text.
Supervision kind (supervision_kind in meta): full_interleaved on every row: the upstream trace itself interleaves text and target images (drawn or rendered states on most sources; the source note above says which), and the read-back comes after the target image it reads. In this release the text was rewritten by a model shown every image of the row (see the r4 section above).
Upstream: multimodal-reasoning-lab/Zebra-CoT. Licence: cc-by-nc-4.0. Changes from upstream: every image is decoded, re-encoded as JPEG and downscaled (the input to 512x288; targets to 911x512, or 448 or 384 px wide on rows flagged S9.px=*); the image placeholders in the question and the reasoning were removed and the sentence around each repaired, sometimes by inserting words such as 'the image'; the reasoning is split into one thought per target; the ThinkMorph system prompt is prepended to the question; rows were filtered and balanced as described below.
Known issues
Rows with a measured per-row problem are listed in reports/known_issues/, one TSV per issue (a # <description> line, then row_uid<TAB>split<TAB>detail lines), so they can be filtered out. They are still in this release: no row was removed for these issues.
| issue | rows | train | validation | what | how it was found | file |
|---|---|---|---|---|---|---|
target_repeats_earlier_target |
665 | 644 | 21 | A target is byte-identical to an earlier target of the same row (a 'remove 0 X objects' step, or objects added and then removed), so the model is trained to redraw an image it has already drawn; S8 compares targets only with the input image. | sha1 of target j equals the sha1 of some target i < j of the same row; measured 2026-10-03; known_issues.py sha1 b2249331; export b3fa1b9 rows 14a4fb2c/18d9ad8f | reports/known_issues/target_repeats_earlier_target.tsv |
readback_states_no_count |
5 | 5 | 0 | The final thought never writes the answer's number: it promises to count, lists the objects without a total, or gives only an intermediate or partial number (a few say it without a numeral, e.g. 'only the purple capsule remains' for 1), so the number exists only in <answer> (read-backs that state a different count are listed in readback_states_other_count instead). | neither the digits of <answer> nor its number word (zero/no/none for 0; for 1, 'the only <singular noun>' but not 'a', 'an' or 'single') appears as a word in the read-back, after removing upstream's 'THOUGHT k:' labels; rows of readback_states_other_count excluded; measured 2026-10-03; known_issues.py sha1 b2249331; export b3fa1b9 rows 14a4fb2c/18d9ad8f | reports/known_issues/readback_states_no_count.tsv |
To leave the listed rows out (the snippet in the loader section downloads reports/known_issues/ with the data):
import glob, os
root = "<root>/zebra_multihop"
drop = {line.split("\t")[0] for f in glob.glob(os.path.join(root, "reports/known_issues/*.tsv"))
for line in open(f) if line.strip() and not line.startswith(("#", "row_uid\t"))}
# keep a row when its row_uid (a column of train/, meta/ and preview/) is not in drop
Size
| split | rows | target image slots | distinct target images |
|---|---|---|---|
| train | 2,865 | 10,556 | 9,850 |
| validation | 89 | 314 | 289 |
A slot is one target position in one row. No target image here is shared between rows: every repeat is inside one row, a target identical to an earlier target of the same row (after a step that leaves the scene unchanged, such as 'remove 0 RED objects'; see Known issues).
| task | train | validation |
|---|---|---|
| multihop_counting | 2,865 | 89 |
Input images per row: 1. Target images per row (the images the model is trained to generate): 2 to 5.
Image corpus (source_scene_corpus): procedural_grid 2,954. procedural_grid is this pipeline's label for any procedurally generated synthetic image (2D grid puzzles and simple 3D renders of primitive objects alike); the source note above says what the images show.
Row format
One row is: input image(s) and a question, then K rounds of thought → target image (the target is the source's own ground-truth image, which the model is trained to generate), then a final thought (normally a read-back of the last target; where a source's final thought is something else, or often leaves out the answer, the source note or Known issues says so) and the answer; here K is 2 to 5. In the train config:
image_list list<binary> inputs first, then the K target images in order
num_input_images int64 how many of image_list are inputs
instruction_list list<string> one element: system prompt + question
output_text_list list<string> K+1 elements:
[0] <think>plan 1</think><image_start>
[j] <image_end><think>plan j+1</think><image_start>
[K] <image_end><think>read-back</think><answer>answer</answer>
row_uid string join key to `meta` and `preview`
<answer> holds exactly meta.answer_value (also the answer column of preview) on every row: score model output against that string.
The system prompt is ThinkMorph's VLM_THINK_SYSTEM_PROMPT from its inferencer.py, verbatim (GEN_THINK_SYSTEM_PROMPT there has the same text), including its leading and trailing newline. The markers are plain strings, not tokenizer special tokens; the prompt writes </image_end> and the data writes <image_end>, exactly as the ThinkMorph-7B checkpoint was trained.
preview shows the same rows with one column per slot: input_image_i for the inputs; for each of the K = num_steps rounds, the plan thought_j and its target target_image_j, both empty for K <= j < 5; and the read-back in thought_5 on every row.
meta holds the per-row sidecar: task, scene_id and geometry_uid (the scene and geometry keys; the split key is named in the split paragraph below), trajectory_id (a camera-path or sample label, empty where the source has none), num_steps, num_input_images, answer_type, answer_value, majority_class_rate, target_image_kind, target_px, est_tokens, licence, split (train / validation, the Hub split names), supervision_kind (full_interleaved / visual_aux / visual_only) and filter_flags. majority_class_rate is the share of the task's most frequent answer_value among its training rows: it measures answer skew and is not a guessing baseline (where a task mixes question types or each row has its own options it can be far below chance); compare scores with the text-only baselines below.
Per-row license in meta: cc-by-nc-4.0 2,954.
Flags on released rows (filter_flags in meta and preview, comma-separated):
| flag | rows | meaning |
|---|---|---|
S5.replay_unsupported |
2,954 | no solver re-derives this task's answer from the trace, so S5 did not replay it |
S9.px=384 |
707 | targets stored at this long-edge size (px) because the row was over the token budget at full size; the trainer's VAE transform (short edge at least 512) scales them back up, so these targets are trained as upsampled, softer images |
S8.phash_near_but_distinct |
282 | a target's perceptual hash is within 6 bits of an input image's, but its pixels differ, so it is not a copy; kept |
S14.sampled_qa |
194 | chosen for the S14 human spot-check (reports/s14_sample.tsv) |
Training with a BAGEL-family loader
Every row here has one input image (num_input_images is 1), so the stock ThinkMorph UnifiedEditIterableDataset (https://github.com/ThinkMorph/ThinkMorph: image_list[0] as input, image_list[j+1] after output_text_list[j]) and the IPT release's version (which reads num_input_images) both read it as intended. Mixed with a source whose rows have more than one input image, only a loader that reads num_input_images is correct.
The stock BAGEL edit loader (ByteDance-Seed/Bagel) cannot train these rows: it never reads output_text_list and expects each instruction_list element to be a list of paraphrases.
parquet_info.json keys each training chunk as <source>/<split>/<file>, here zebra_multihop/train/chunk_00000.parquet, with row-group counts read from the parquet footers. The loader matches a chunk only when its key equals the path it builds, os.path.join(data_dir, file), and skips a chunk with no key without a warning: a source that is alone in its group then fails with IndexError: list index out of range, and in a mixed group it adds no rows. Download into a directory named after the source, not after the repository:
from huggingface_hub import snapshot_download
snapshot_download("yrlyrl/spatial-mmcot-zebra_multihop", repo_type="dataset", local_dir="<root>/zebra_multihop",
allow_patterns=["train/*", "validation/*", "parquet_info.json", "reports/known_issues/*"])
Then either run from <root> with data_dir: zebra_multihop/train and parquet_info_path: zebra_multihop/parquet_info.json, or rebuild the index with absolute keys and use an absolute data_dir:
import json, os
root = "/abs/path/to/root" # the directory that holds zebra_multihop/
info = json.load(open(os.path.join(root, "zebra_multihop", "parquet_info.json")))
info = {os.path.join(root, k): v for k, v in info.items()}
json.dump(info, open(os.path.join(root, "zebra_multihop", "parquet_info_abs.json"), "w"))
# data_dir = os.path.join(root, "zebra_multihop", "train") (spelled exactly so, no trailing slash)
# parquet_info_path = os.path.join(root, "zebra_multihop", "parquet_info_abs.json")
The Hugging Face cache (.../snapshots/<hash>/train/) or a folder named spatial-mmcot-zebra_multihop matches no key.
num_used_data counts chunk files, not rows: the loader repeats this source's file list up to that number, lists every (file, row group) pair, and deals whole row groups out, floor(R / world_size) to each rank and floor(that / num_workers) to each DataLoader worker. The remainder is never read. This source has 1 training chunk file holding 23 row groups of up to 128 rows, so keep num_used_data large, e.g. the 128 of ThinkMorph's interleaved_reasoning.yaml (upstream's example.yaml asks for more than GPUs x workers); every row group is then read. Set to 1 and alone in its group on 8 GPUs with 4 workers, it gives every DataLoader worker an empty list, and the iterator then loops forever printing repeat without yielding a row. In a run that mixes sources, give each source the same multiple of its own training chunk-file count, e.g. 128 per file (128 here): the file list is repeated up to num_used_data entries, so a flat 128 for every source would read a two-file source's rows half as often as a one-file source's.
How the rows were chosen
| stage | rows |
|---|---|
| upstream rows read | 10,000 |
refused before conversion (S0raw; each reason is in the table below) |
3,203 |
| quarantined at S4c (an automatic check could not match the read-back's conclusion to the label) | 92 |
| dropped at S5 (the text contains a phrase from S5's self-contradiction list, e.g. 'does not make sense', 'discrepancy', 'there must be a mistake'; a keyword match, not a comparison with the images, so it also removes some sound rows) | 17 |
| dropped at S8 (a target was removed as a copy of an input, as transparent or as too small, and the row had no target left or was a multi-step chain that cannot lose a state) | 918 |
| dropped at S9 (over the token budget or too many target images) | 2,057 |
| refused by the final structural check (S0, after S9) | 33 |
| after conversion and per-row filters | 3,680 |
| removed by S10 (none) | 0 |
| removed by answer-prior balancing (S13) | 614 |
removed for training after the r4 rewrite (subset_dropped.jsonl) |
112 |
| released | 2,954 |
Every removed row has one line, with its reason, in reports/:
| file | step | reason (the line's flag, or the field shown) |
rows |
|---|---|---|---|
build/dropped.jsonl |
S0 | S0.empty_readback |
33 |
build/dropped.jsonl |
S0raw | S0.non_integer_answer |
2,061 |
build/dropped.jsonl |
S0raw | S0.ops_not_stated |
1,142 |
build/dropped.jsonl |
S5 | S5.self_contradiction |
17 |
build/dropped.jsonl |
S8 | S8.chain_broken |
918 |
build/dropped.jsonl |
S9 | S9.needs_truncation |
2,057 |
build/quarantine.jsonl |
S4c | S4c.cot_label_conflict |
92 |
s13_dropped.jsonl |
S13 | step: answer |
614 |
subset_dropped.jsonl |
r4 | reason: over_budget |
77 |
subset_dropped.jsonl |
r4 | reason: label_doubt |
35 |
Every line of s13_dropped.jsonl has reason: prior_downsample; step names the balancing pass that removed it, and split is written train or val (the Hub's validation).
S0raw lines in build/dropped.jsonl were refused before a release row existed, so their row_uid field holds the converter's key for the upstream record, not a 16-hex row_uid; lines from later steps carry the row_uid the row had. No removed row appears in meta or preview. For this source the key is built from upstream fields that repeat across rows (for most sources a hash of the question text), so it is not unique: the 2,061 S0.non_integer_answer lines carry 2,059 distinct keys. Those rows are counted with their reason but cannot be traced to individual upstream rows.
92 rows were quarantined rather than dropped (S4c): an automatic check could not match the read-back's stated conclusion to the stored label. They were not reviewed by hand, and some are phrasing mismatches rather than wrong labels.
S8.k_zero and S8.chain_broken name what happened to the row, not which check removed the image; the lines in this build do not record whether it was the size, transparency or copy check.
Per-step counters of the conversion
10,000 upstream rows were read; S0raw refused 3,203 before a row existed and passed 6,797 to the first step. S0 runs once more, last, on the final bytes. The reason for every refused, dropped or quarantined row is in the files above.
| step | in | out | dropped | quarantined | rejected | repaired |
|---|---|---|---|---|---|---|
| S0raw (refused before conversion) | 10,000 | 6,797 | 0 | 0 | 3,203 | 0 |
| S4 | 6,797 | 6,797 | 0 | 0 | 0 | 0 |
| S4c | 6,797 | 6,705 | 0 | 92 | 0 | 0 |
| S5 | 6,705 | 6,688 | 17 | 0 | 0 | 0 |
| S8 | 6,688 | 5,770 | 918 | 0 | 0 | 0 |
| S9 | 5,770 | 3,713 | 2,057 | 0 | 0 | 0 |
| S0 (final structural check, after S9) | 3,713 | 3,680 | 33 | 0 | 0 | 0 |
The train/validation split keeps rows sharing a scene_id in meta on one side, and the assignment is frozen (splits/ in the summary repository). scene_id is the md5 of the upstream input image (problem_image_1), so a key is one input image; geometry_uid repeats it and trajectory_id is empty. S12 saw 3,680 rows under 3,680 keys, one row per key, so the split is in effect per row. No validation input image has the content of a training input image. The S12 run did not record whether its pixel-level near-copy test ran for this source, so near-copies are not ruled out.
Answer-prior balancing (S13)
Each (task, split) group is checked separately. An answer is the answer value compared as lower-cased text without a trailing full stop, with 'farther' read as 'further' and 'nearer' as 'closer' (for multiple choice, the option text, not the letter; where the candidates are drawn in the image, as in zebra_jigsaw and zebra_tetris, the answer is the letter itself). An answer is real when it holds at least 5 rows and 2% of the group; k is the number of real answers. Answer step: the target is max(30%, 1/k) when k >= 2, and max(30%, 1/d) over the d distinct answers when k = 1; a validation group uses the larger of its own target and its task's train target. A group is cut only when k >= 1 and its most common answer holds more than the target plus 5 percentage points; every answer is then capped at one common count, chosen so that none exceeds the target, and smaller answers keep all their rows. At the answer step, a group at or below that trigger, or with no real answer (k = 0), is left as it is, so its most common answer can hold up to the target plus 5 percentage points. A task whose train group has exactly two real answers is instead cut, in every split, so that its two largest answers have equal counts, with no trigger. Rank and label steps: then, in a group where every option value of every row is a number, the rank of the correct option among the sorted values, and after it, in a group where every trained answer is an option label, the label, are each capped by the same cut-and-trigger rule on their own counts (own target, validation included): capped, never evened out, so two labels are cut only when one exceeds 55%, and then only down to 50%. These steps can also cut groups the answer step left whole, including k = 0 groups, and can raise an answer's final share above its target; the run fails if a real answer ends above the target plus 5 percentage points. A train group of at least 20 rows in which one answer holds 90% or more fails the run. PET (exact_cells_pet) instead cuts each (question type x turn direction) cell to equal counts of its two answers; a PET cell that shows only one answer is removed.
| task | split | pass | rule | rows in → out | real answers k | target | largest share, before → after | cut |
|---|---|---|---|---|---|---|---|---|
| multihop_counting | train | answer | cap30[canon] |
3,574 → 2,972 | 8 | 30.0% | 41.8% → 30.0% | yes |
| multihop_counting | validation | answer | cap30[canon] |
106 → 94 | 7 | 30.0% | 37.7% → 29.8% | yes |
S13 removed 614 rows from this source.
Text-only baselines
Accuracy of guessers that never see an image. For each task the released training rows are split into two fixed halves by a hash of row_uid; each guesser is fitted on one half and scored once on the other (one held-out half, not cross-validation; eval rows below). The reference is chance (the mean of 1 / number of options) where every row is multiple choice, and otherwise the eval-half accuracy of always giving the answer most common in the fit half (when a task's top answers are nearly tied, this need not be the task's most common answer; the line after the table gives that answer's validation score). Accuracies are recounted from the stored rates and eval rows, so they are exact. A task is flagged when a text-only guesser beats its reference by more than 0.15 (for a free-form task, a guesser other than the most common answer). A flagged task can be partly answered from the text alone; an unflagged task passed only these probes, which do not prove the text carries no answer. Report scores on every task next to this baseline.
Guessers: keywords: the most common answer per set of spatial words in the question; majority: the answer most common in the fit half; template: the most common answer per question wording (numbers masked, object names kept).
| task | best text-only guesser | accuracy | reference | margin | eval rows | flagged |
|---|---|---|---|---|---|---|
| multihop_counting | template |
0.294 | 0.294 (majority) | +0.000 | 1,498 | no |
Always giving the most common training answer, scored on the validation split (the constant baseline to compare validation scores with): multihop_counting: always answering 1 (30.2% of training rows) scores 0.303 (27/89).
Spot-check (S14)
Pending. The S14 rows are chosen and flagged S14.sampled_qa in meta and preview; the human pass over them has not been signed off yet.
Citation
Zebra-CoT is by Ang Li, Charles Wang, Deqing Fu, Kaiyu Yue, Zikui Cai, Wang Bill Zhu, Ollie Liu, Peng Guo, Willie Neiswanger, Furong Huang, Tom Goldstein and Micah Goldblum, released under CC BY-NC 4.0; this repository is a converted subset of it (the changes are listed under the licence line above). Its card asks users to cite:
@inproceedings{li2026zebracot,
title={Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning},
author={Ang Li and Charles Wang and Deqing Fu and Kaiyu Yue and Zikui Cai and Wang Bill Zhu and Ollie Liu and Peng Guo and Willie Neiswanger and Furong Huang and Tom Goldstein and Micah Goldblum},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=c6XIVI3TiQ}
}
Provenance
The release files were written by our conversion code (the code repository is not public yet), scripts/convert/export.py at commit b3fa1b993ca9, from build zebra_multihop_r3. The build was made by scripts/convert/run_source.py from the same repository at commit 76c78bd817c2. S10, S12 and S13 ran before the export; reports/export_manifest.json pins every input the export read by SHA-1 (build_manifest_sha1, s10_keep_sha1, s12_assignments_sha1, s13_balanced_keep_sha1).
Every row removed between upstream and this release has one line, with its reason, in reports/: build/dropped.jsonl (rows refused before conversion or dropped by a conversion step); build/quarantine.jsonl (rows set aside by S4c because an automatic check could not match the read-back's conclusion to the label); s10_dropped.jsonl (duplicates removed by S10); s10_label_conflicts.jsonl (rows S10 withheld because another row asks the identical question, options in the same order, of the same images with a different answer); s13_dropped.jsonl (rows removed by answer-prior balancing). known_issues/ lists rows with a measured problem (see Known issues); reports/ also holds the build manifest (absolute paths cut to basenames) and counters, the S14 sample list (s14_sample.tsv: row_uid, task, split) and export_manifest.json. Part of yrlyrl/spatial-mmcot.
- Downloads last month
- 454