Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringclasses
10 values
task
stringclasses
1 value
question
stringclasses
10 values
answer
stringclasses
1 value
input_images
images listlengths
1
1
reasoning_thought_0
stringclasses
10 values
reasoning_image_0
imagewidth (px)
341
512
reasoning_thought_1
stringclasses
10 values
e02074d04c24bddfb8d025695f852825282f83146642f2aeb6ae08206332e507
bbox
What is the main color of the vest worn by the person standing near the stage on the left side of the image? (A) Bright yellow-green (B) Dark blue (C) Red (D) White Answer with the option's letter from the given choices directly.
A
To answer the question, first look at the left side of the image where the stage structure is located. The performer in red is standing on a narrow platform along the front of the stage, and the question asks about another person standing near the stage on that same left side. Scan downward and slightly behind the perf...
The highlighted region isolates the vest worn by the person standing near the stage on the left. Its color is a vivid fluorescent yellow-green, the kind commonly used for security or event staff, and it clearly does not appear dark blue, red, or white. The strong neon tone makes the correct choice A.
be44205d9e0c7b3e365d1a9e494587ef0229e804eda509f2fe09e2f6ceecd458
bbox
What main color is the lower part of the poster? (A) Green (B) Blue (C) Red (D) Yellow Answer with the option's letter from the given choices directly.
A
To answer the question, first locate the poster in the scene rather than focusing on the large red-and-white church building. Look toward the lower middle part of the image, near the walkway in front of the small red-and-white guardhouse-like structure. There is a tall, narrow signboard or poster standing upright on th...
The highlighted region isolates the tall poster standing in front of the small red-and-white structure near the lower center of the image. Looking closely at the poster, its upper area is mostly light or white, while the bottom section contains a distinctly green panel. Because the question asks for the main color of t...
34e25eab9e63ac38f5187cd133a9024107e84aba57555d95aa61060ed532cbff
bbox
What is the color of the van? (A) White (B) Blue (C) Red (D) Yellow Answer with the option's letter from the given choices directly.
A
To answer the question, first look across the open plaza and identify any vehicles, since the prompt asks specifically about a van. Most of the image is dominated by the large stone arch in the center and long yellow buildings on both sides, with many pedestrians scattered in the foreground. There are no vehicles in th...
The highlighted region isolates the small van parked along the right side of the square, in front of the yellow building and beneath the arcade. With the surrounding pedestrians and architecture removed from attention, it is easier to judge the vehicle’s body color. The van’s exterior appears bright and light-toned rat...
f14088203e55ee22ee028834aa0c45b9b1f815787fd692068ed5e03dd13d92ab
bbox
What is the main color of the bag carried by the person near the right side of the image? (A) Black (B) Red (C) Blue (D) Brown Answer with the option's letter from the given choices directly.
A
To answer the question, first look toward the right side of the street scene, near the white building with the dark red door. There is a person standing on the sidewalk beside the parked silver SUV. Since the question asks about the bag carried by the person near the right side of the image, focus on that individual ra...
The highlighted region isolates the bag being carried by the person standing next to the silver SUV on the right. Inside the box, the bag appears very dark, with no strong red, blue, or brown tones visible. Its overall appearance is black, especially compared with the person's lighter clothing and the pale wall behind ...
08ed58f263a5280488b734f59760ddc8cbb1be69b73166f67b4b4e0c5a71abc2
bbox
What is the primary color of the wheel on the green trash bin? (A) Black (B) Silver (C) Red (D) Yellow Answer with the option's letter from the given choices directly.
A
To answer the question, first locate the green trash bin in the scene. It is near the lower-right area of the image, standing on the edge of the pavement beside the crowd. Once you find the bin, look down toward its base because the wheel is attached near the bottom. The wheel is small and partially tucked behind the b...
The highlighted region isolates the wheel at the bottom-right side of the green trash bin, making it easier to judge its color without distraction from the pavement or shadows nearby. Looking closely at the wheel inside the marked area, it appears very dark rather than metallic, bright red, or yellow. This matches blac...
4f2ea8461c299aca787be80d917f55f1d75efc765156fa2ad232d49fc2d28eeb
bbox
What is the primary color of the bus on the street near the bottom left of the image? (A) White (B) Red (C) Blue (D) Yellow Answer with the option's letter from the given choices directly.
A
To answer the question, first look toward the lower-left portion of the aerial city scene, where a curved street runs beside the canal and near a cluster of trees and buildings. The prompt specifically asks about a bus on the street near the bottom left, so scan that roadway for a long, rectangular vehicle larger than ...
The highlighted vehicle is clearly the bus referenced in the question. It sits on the curved road in the lower-left area of the image, and its body appears predominantly light-colored. While there are darker windows and some red/orange accents near the lower section, the main exterior color covering most of the bus is ...
ac8122aa66e28a41806ebb4461f55b85a4d199b9d78df6089087af2c6a7b3697
bbox
What material are the side panels of the trash can primarily made of? (A) Wood (B) Glass (C) Concrete (D) Metal mesh Answer with the option's letter from the given choices directly.
A
To answer the question, first look for the trash can in the large outdoor plaza rather than the buildings, trees, or stairways. The trash can is small compared with the rest of the scene, so focus on the lower middle area of the image near the walkway and the decorative wire sculpture. There is a person sitting near th...
With the trash can highlighted, you can now judge the material of its side panels by their color and texture. The sides appear as brown slatted panels with the look of wooden boards rather than transparent glass, solid gray concrete, or an open crisscross metal mesh. The visible paneling has a natural wood tone and seg...
f7c848db37f3fd828d256a56eb00aac5dd6931344879a7866d117b961ccec03a
bbox
What activity is prohibited according to the logo on the street sign next to the statue? (A) Riding bicycles (B) Parking cars (C) Walking dogs (D) Crossing the street Answer with the option's letter from the given choices directly.
A
Start by locating the statue mentioned in the question: it is the large dark sculpture in the center showing a bear standing on its hind legs and reaching toward a tree. The question asks about the street sign next to the statue, so look just to the right side of the statue rather than at the red do-not-enter sign on t...
With the sign now highlighted, you can focus on the symbol inside it. The logo shows a bicycle icon inside a prohibition marking, indicating that bicycles are not allowed. This matches the option about riding bicycles rather than parking cars, walking dogs, or crossing the street. A
0612ff29f98c8cb2f4c8ab5159bda6c92f18523398ac5bdaaca86872b3d1f461
bbox
What type of numerals are used on the clock on the building? (A) Roman numerals (B) Arabic numerals (C) Dots instead of numbers (D) No numerals, only tick marks Answer with the option's letter from the given choices directly.
A
To answer the question, first look for any clock attached to the buildings along the crowded street. Most of the scene is filled with pedestrians and storefront signs, so the important detail is higher up on the buildings rather than at street level. Scan the upper middle part of the image, where the rooftops overlap. ...
The highlighted region isolates the visible part of the clock face, making it easier to inspect the markings around the dial. In that narrow section, the numbers are shown as letter-like symbols rather than standard 1, 2, 3 digits. The visible characters match Roman numeral styling used on traditional clocks. Since the...
170386074442f38422640ccc8acf89fda62e88fa5686169686030b557a7ebedf
bbox
What color is the rear light on the scooter? (A) Red (B) Blue (C) Green (D) Yellow Answer with the option's letter from the given choices directly.
A
Start by locating the scooter in the street scene. The road runs through the left half of the image, while pedestrians walk along the sidewalk on the right. Near the middle-left area of the road, just behind the parked or slow-moving cars, there is a small dark scooter with a rider on it. Because the question asks abou...
The highlighted region isolates the rear section of the scooter, making the tail light easier to inspect. Inside that boxed area, the light on the scooter’s back is visibly red, standing out against the darker body of the scooter and the gray road surface. Since the question asks for the color of the rear light on the ...

Instruct-180K: Hard Visual Search, Jigsaw and Depth Comparison

This dataset is the harder end of three task groups from Instruct-180K, the Stage 2 data of PRIMER that trains ThinkMorph-SSL-Instruct. Each trace has a question, input images, textual reasoning with intermediate images, and an answer.

Counts. 17,895 traces in five task folders, with 37,427 images (50.4 GB).

Group Task (folder) Operation Rows Of in Instruct-180K
Visual search bbox box drawn on the image 3,000 14,686
Visual search crop zoomed crop 3,000 7,479
Jigsaw jigsaw_single_image pieces of one image reassembled 6,000 19,935
Depth comparison depth_map depth map, then which object is closer, farther or in front 3,858 6,849
Depth comparison marked_depth_map query points marked, then the depth map, then which point is closer, farther or in front 2,037 2,726
Total 17,895

How the rows were chosen

Each task keeps its hardest rows by a simple, measurable proxy, and ties are broken by sha256("vsj-v1:" + id) so the cut is reproducible.

  • bbox. Only high-resolution inputs (at least 1 MP) are eligible, which leaves out the low-resolution, mostly open-ended Visual-CoT rows. The drawn box is recovered as the region where the boxed image differs from the input, and the 3,000 rows with the smallest box relative to the image are kept.
  • crop. The zoomed crop is an unscaled region of the input, and the 3,000 rows with the smallest crop relative to the image are kept.
  • jigsaw_single_image. Two-piece puzzles are left out. All 3,994 2×2 puzzles (24 arrangements) are kept, plus 2,006 three-piece puzzles (1×3 and 3×1).
  • Depth comparison. Every depth row that compares the depth or distance of two or more points or objects is kept, because that is about 6k rows in total. Rows that ask for an absolute distance, a point location, a size or a camera rotation are left out. The rule is keyword-based and was checked by reading samples, so a few borderline rows may remain.
  • MANIFEST.json records the source manifest's SHA256 and, per task, the pool size, the rule and the score cutoff in cut_from.

Every kept row is the same line, byte for byte, as in Instruct-180K at revision 80dd63f8eb02337b3cd46724cef8f40eaaa12a26, in the same order. Every image keeps its content-hashed name and bytes.

Files

Path What it is
shards/SHARDS.json, shards/data-*.tar The release: MANIFEST.json, the five task files and all images, in tar shards of about 1 GiB
preview/*.parquet The Dataset Viewer preview: one config per task, 10 rows each, images downscaled to at most 512 px

Download and restore

hf download linjieli222/Instruct-180K-Search-Jigsaw-Depth-Hard --repo-type dataset --include "shards/*" --local-dir .
python scripts/shard_data_release.py restore shards/SHARDS.json shards DATA_ROOT/instruct-180k-search-jigsaw-depth-hard

restore checks every shard's size and SHA256 before extracting, then checks every file against MANIFEST.json. A restored release holds one folder per task with train.jsonl, one shared images/ folder and MANIFEST.json. Each task folder is a dataset folder for the BAGEL loader. The scripts are in the PRIMER code release.

Row format

{"id": "<sha256>", "task": "bbox", "question": "...", "options": "...",
 "images": ["../images/<sha256>.jpg"], "reasoning_thought_0": "...",
 "reasoning_image_0": "../images/<sha256>.jpg", "reasoning_thought_1": "...", "answer": "..."}
  • The supervision text and the image bytes are exactly as in the training copy.
  • Image paths are relative to the row's file.
  • Empty question, options and answer fields are left out, and the BAGEL loader reads a missing field as empty.

Source terms

Each source keeps its own terms, and the image release has no blanket license. The visual-search images come from Visual-CoT and visual-search image pools, the jigsaw images from jigsaw image pools, and the depth images from RefSpatial. See the Instruct-180K card for provenance details.

Downloads last month
148