Datasets:
id int64 | original_id int64 | context string | n_images int64 | placed_depth float64 | target_index int64 | target_category string | images_list list | context_length int64 | needle_caption string | mcq_options dict | mcq_answer string |
|---|---|---|---|---|---|---|---|---|---|---|---|
0 | 163 | Underneath the paint and the gunge
With all components removed from the chassis and the black sealant (gunge) scraped away the poor state of our Ensign was now clear for all to see. Severe rust, corrosion, damage and bad repairs left us in no doubt that this restoration project was going to be a real challenge.
So bad... | 5 | 0.2 | 3 | donut | [
"train2017/000000421681.jpg",
"val2017/000000389804.jpg",
"train2017/000000287213.jpg",
"train2017/000000312139.jpg",
"train2017/000000579043.jpg"
] | 9,137 | A table topped with frosting covered donuts and diet coke. | {
"A": "Donuts in bins at a display with labels in a foreign language.",
"B": "Donuts on a donut machine being made ",
"C": "A table topped with frosting covered donuts and diet coke.",
"D": "A man holding two glazed donuts in his hands."
} | C |
1 | 195 | We spent a week at this circus-themed hotel and loved every bit of its elegance and quirkiness. Top to bottom, the 25hours Hotel impressed us with its ground floor breakfast buffet and top floor bar/lounge. I highly recommend heading up for a drink of local brewed beer and panoramic view of the Viennese skyline. I hear... | 5 | 0.89 | 3 | orange | [
"train2017/000000447927.jpg",
"train2017/000000354321.jpg",
"train2017/000000124236.jpg",
"train2017/000000023400.jpg",
"train2017/000000182048.jpg"
] | 9,285 | A number of ripe oranges inside of a mesh container. | {
"A": "An orange cut pieces and knife are kept on a board.",
"B": "A plate has a dessert and orange slices on it.",
"C": "Butterflies are sitting on the many orange slices.",
"D": "A number of ripe oranges inside of a mesh container."
} | D |
2 | 236 | Set on a small cove in Asinara National Park, this white-sand beach is about as remote as it gets in Sardinia. The park, which derives its name from the local herd of albino donkeys, encompasses the entirety of uninhabited Isola dell’Asinara — Sardinia’s second-largest island. Before its designation as a national park,... | 5 | 0.82 | 1 | keyboard | [
"train2017/000000465267.jpg",
"train2017/000000474446.jpg",
"train2017/000000477882.jpg",
"train2017/000000548661.jpg",
"train2017/000000281790.jpg"
] | 11,074 | A dachshund dog standing on a desk next to a stuffed animal penguin and a laptop computer and keyboard. | {
"A": "a desk-top cluttered with phone, a keyboard and papers ",
"B": "A dachshund dog standing on a desk next to a stuffed animal penguin and a laptop computer and keyboard.",
"C": "Two lap tops are open on a table near a phone and a keyboard.",
"D": "A computer keyboard sitting on top of a wooden table."
} | B |
3 | 173 | – The actress noted that there is a lot happening online and as such, they should occasionally log off
Celebrated Nollywood actress Omotola Jalade-Ekeinde recently took to her page on Twitter with some words of advice for her fans and followers.
The movie star in her tweet submitted that there is a lot of pressure on... | 5 | 0.48 | 2 | umbrella | [
"train2017/000000208612.jpg",
"train2017/000000127076.jpg",
"train2017/000000369074.jpg",
"train2017/000000055184.jpg",
"train2017/000000135258.jpg"
] | 9,565 | A man is holding an umbrella while riding an elephant | {
"A": "Four women stand on the beach with their umbrellas.",
"B": "A man is holding an umbrella while riding an elephant",
"C": "A person on the boardwalk with a multicolored umbrella. ",
"D": "A area of concrete with several blue umbrellas laying on it."
} | B |
4 | 223 | The underwriters argued, with good evidence, that the entire voyage had been badly mismanaged. They also argued that the crew should be prosecuted for murder. Granville Sharp, an active campaigner against the slave trade, probably suggested the latter strategy. A former slave, Olaudah Equiano, had brought Sharp's atten... | 5 | 0.48 | 4 | bed | [
"train2017/000000012313.jpg",
"train2017/000000289061.jpg",
"train2017/000000189888.jpg",
"val2017/000000394677.jpg",
"train2017/000000412326.jpg"
] | 10,706 | A bed with a nightstand that has tissues and other things on top | {
"A": "BRIGHTLY LIT BEDROOM WITH TWO SIDE TABLE WITH LAMPS",
"B": "A teddy bear tucked into a bed that's been made",
"C": "Two monogrammed pillows are stacked on a bed.",
"D": "A bed with a nightstand that has tissues and other things on top"
} | D |
5 | 157 | This corner softness is unrelated to depth of field; in fact, in the above example, the left and right photos depict objects which were about the same distance from the camera. Instead, this corner softness is entirely caused by how a lens is designed. For a given lens, the only way to reduce this edge softness is to u... | 5 | 0.16 | 0 | baseball bat | [
"train2017/000000415016.jpg",
"train2017/000000451574.jpg",
"train2017/000000013565.jpg",
"train2017/000000154610.jpg",
"train2017/000000282377.jpg"
] | 9,296 | A young boy holding a baseball bat on a field. | {
"A": "A young man poses with a ball bat.",
"B": "A young child in a khaki pants and a hitting helmet, holding a small baseball bat while standing in the grass.",
"C": "A baseball player with a bat in his hands on a field. ",
"D": "A young boy holding a baseball bat on a field."
} | D |
6 | 165 | "Read Contemporary French Philosophy Modernity And The Persistence Of The Subject\n\nRead Contempora(...TRUNCATED) | 5 | 0.22 | 2 | bed | ["train2017/000000083561.jpg","train2017/000000453376.jpg","train2017/000000552217.jpg","train2017/0(...TRUNCATED) | 10,199 | A person in bed turning off a clock radio. | {"A":"An empty bed in a bedroom with light coming out from underneath it.","B":"A person in bed turn(...TRUNCATED) | B |
7 | 170 | "Novosibirsk (the heart of Ural) named New Year's Capital of Russia and Kemerovo Region the Partner (...TRUNCATED) | 5 | 0.41 | 3 | banana | ["train2017/000000317234.jpg","train2017/000000421069.jpg","train2017/000000061144.jpg","train2017/0(...TRUNCATED) | 9,115 | A close-up of the monkey sitting down eating a banana. | {"A":"A close-up of the monkey sitting down eating a banana.","B":"Green bananas on tree outside wit(...TRUNCATED) | A |
8 | 187 | "Novosibirsk (the heart of Ural) named New Year's Capital of Russia and Kemerovo Region the Partner (...TRUNCATED) | 5 | 0.72 | 1 | horse | ["train2017/000000378765.jpg","train2017/000000581813.jpg","train2017/000000503534.jpg","train2017/0(...TRUNCATED) | 9,115 | The horses are grazing in the grassy field. | {"A":"A man drives a horse cart with a cow in the back","B":"The horses are grazing in the grassy fi(...TRUNCATED) | B |
9 | 192 | "Thankfully those concerns are incidental ones, not a constant pain in the arse, and most of my time(...TRUNCATED) | 5 | 0.82 | 1 | elephant | ["train2017/000000569769.jpg","train2017/000000449718.jpg","train2017/000000037429.jpg","train2017/0(...TRUNCATED) | 9,510 | an elephant is in its pen and in some water | {"A":"A pair of elephants standing in their natural habitat.","B":"Three elephants standing near eac(...TRUNCATED) | C |
End of preview. Expand in Data Studio
ICIR-MCQ
The controlled probe used in "Heads That Write, Not Just Point: Image Retrieval Heads in
Vision-Language Models" (NeurIPS 2026) to discover image retrieval heads. The 6,000 samples
used in the paper are in coco_retrieval_n5n6_mcq.jsonl, exactly as used.
- Code (head extraction, pointer and writer rankings, ablations): https://github.com/parkquasar/Image-Retrieval-Heads
- Paper: NeurIPS 2026 (OpenReview link to be added when the proceedings are published)
from datasets import load_dataset
ds = load_dataset("jerryray/ICIR-MCQ", split="train") # 6,000 rows
Contents
coco_retrieval_n5n6_mcq.jsonl: 6,000 rows (3,000 with five candidate images, then 3,000 with six)build_coco_retrieval_dataset_mcq.py: the builder that produced itSHA256SUMS
Images are not included. Download MS-COCO 2017 train/val; paths in the jsonl are COCO-relative
(train2017/000000421681.jpg).
Sample format
One sample per line. Five or six images are interleaved with long text passages, and the task is to pick, among four captions, the one describing the target image.
| field | meaning |
|---|---|
id |
0 to 5999 |
context |
prompt body; contains n_images occurrences of <image>, filled in order |
n_images |
5 or 6 |
images_list |
COCO paths in <image> order |
target_index |
index of the target image in images_list (0-based) |
target_category |
COCO category named in the question |
mcq_options |
captions A to D: one correct, three captions of other images of the same category |
mcq_answer |
correct letter |
needle_caption |
caption of the target image (equals mcq_options[mcq_answer]) |
placed_depth |
relative position of the target (0 to 1) |
context_length |
text length in tokens |
original_id |
index of the MM-NIAH document the text came from |
The question is appended to context as below; the paper experiments use this format.
Which caption best describes the image above that contains {target_category}? Answer with a single letter (A, B, C, or D).
A) ...
B) ...
C) ...
D) ...
Score by comparing the first letter A to D in the response with mcq_answer.
Statistics
Answers A/B/C/D are balanced (1,500 each) and the target slot is balanced as well. All 33,000 images are distinct (no reuse) across 77 categories. Text length is 9k to 13k tokens, median 11k.
Sources
- Images, captions and categories: MS-COCO 2017. Images are not redistributed; only file names.
- Interleaved text: the retrieval-image documents of MM-NIAH (OpenGVLab, MIT licence), which originate from OBELICS (CC-BY-4.0).
Please keep these attributions if you redistribute the data.
Rebuilding
export COCO_ANN_ROOT=/path/to/COCO/annotations # captions_*2017.json, instances_*2017.json
export COCO_IMG_ROOT=/path/to/COCO/images # train2017/, val2017/
export MM_NIAH_ANN=/path/to/MM-NIAH/mm_niah_val/annotations/retrieval-image.jsonl
python build_coco_retrieval_dataset_mcq.py --output-n5 n5.jsonl --output-n6 n6.jsonl --seed 42
cat n5.jsonl n6.jsonl > coco_retrieval_n5n6_mcq.jsonl
Seed 42 reproduces the shipped file bit for bit. Check integrity with sha256sum -c SHA256SUMS (the data file and the builder are listed; this README is not).
License
The dataset file (captions, questions, options, image file names and the interleaved text) is released under CC BY 4.0. Captions and categories come from MS-COCO 2017 (annotations CC BY 4.0); the interleaved passages come from MM-NIAH (MIT), which draws on OBELICS (CC BY 4.0). COCO images are not redistributed. The builder script is MIT-licensed, as is the accompanying code repository.
Citation
@inproceedings{park2026heads,
title = {Heads That Write, Not Just Point: Image Retrieval Heads in Vision-Language Models},
author = {Park, Junsung and Yu, Sangwon and Kang, Donghun and Oh, Yeongtak and Jeong, Jaewon and Yoo, Gyeongtae and Moon, Han Cheol and Lee, Jungbeom and Yoon, Sungroh},
booktitle = {Advances in Neural Information Processing Systems},
year = {2026}
}
- Downloads last month
- 38