Dataset Viewer
Auto-converted to Parquet Duplicate
id
int64
1
1k
image
imagewidth (px)
142
8.69k
image_name
stringlengths
11
14
prompt
stringclasses
1 value
truth
listlengths
2
13
hallu
listlengths
3
6
1
AMBER_1.jpg
Describe this image.
[ "sky", "forest", "grass", "person", "lake", "mountain", "road" ]
[ "cloud", "sun", "bird", "dog", "flower" ]
2
AMBER_2.jpg
Describe this image.
[ "sky", "man", "ship", "ship", "forest", "bridge", "mountain", "cloud", "lake", "building" ]
[ "plane", "bird", "sun", "paddle", "ground" ]
3
AMBER_3.jpg
Describe this image.
[ "child", "grass", "flower" ]
[ "sky", "cloud", "tree", "house", "bench" ]
4
AMBER_4.jpg
Describe this image.
[ "sky", "cloud", "forest", "lake", "woman", "paddle", "watch" ]
[ "ship", "sun", "ground", "bird", "grass" ]
5
AMBER_5.jpg
Describe this image.
[ "person", "sea", "beach", "child", "toy" ]
[ "ship", "sky", "reef", "seegull", "lounge" ]
6
AMBER_6.jpg
Describe this image.
[ "sky", "mountain", "bicycle", "person", "hat", "guardrail", "sign", "tree", "ground", "road", "cow", "sheep" ]
[ "motorbike", "car", "bird", "sun", "cloud" ]
7
AMBER_7.jpg
Describe this image.
[ "sea", "beach", "bucket", "shovel", "toy", "child" ]
[ "sky", "mountain", "ship", "lounge", "dog" ]
8
AMBER_8.jpg
Describe this image.
[ "sky", "forest", "grass", "duck", "child", "hat", "road", "goose" ]
[ "cloud", "sun", "woman", "dog", "toy", "car" ]
9
AMBER_9.jpg
Describe this image.
[ "football", "man", "grass", "court" ]
[ "goal", "sky", "sun", "ground" ]
10
AMBER_10.jpg
Describe this image.
[ "ball", "man", "hat", "sky", "tie" ]
[ "sun", "bird", "grass", "tree", "woman" ]
11
AMBER_11.jpg
Describe this image.
[ "dog", "beach" ]
[ "sea", "ship", "sky", "person", "toy" ]
12
AMBER_12.jpg
Describe this image.
[ "person", "grass", "rugby", "child" ]
[ "sky", "cloud", "dog", "bench", "house" ]
13
AMBER_13.jpg
Describe this image.
[ "grass", "horse", "woman", "forest" ]
[ "sky", "sun", "ground", "fence", "man" ]
14
AMBER_14.jpg
Describe this image.
[ "sky", "grass", "forest", "lake", "woman", "beach", "child", "ground" ]
[ "fish", "bird", "cloud", "umbrella", "dog" ]
15
AMBER_15.jpg
Describe this image.
[ "desert", "sky", "cloud", "hat", "skate", "man", "mountain" ]
[ "umbrella", "sun", "car", "camel" ]
16
AMBER_16.jpg
Describe this image.
[ "sky", "cloud", "sea", "beach", "dog", "mountain" ]
[ "lounge", "umbrella", "tree", "person", "seegull" ]
17
AMBER_17.jpg
Describe this image.
[ "person", "surfboard", "sea" ]
[ "ship", "sky", "mountain", "beach", "tree" ]
18
AMBER_18.jpg
Describe this image.
[ "sky", "cloud", "man", "mountain", "tree", "ground" ]
[ "camera", "dog", "sun", "tree", "bird" ]
19
AMBER_19.jpg
Describe this image.
[ "tree", "sky", "person", "ground", "lake", "hat", "cloud", "crutch", "sun", "mountain" ]
[ "bird", "flower", "house", "insect", "squirrel" ]
20
AMBER_20.jpg
Describe this image.
[ "sky", "sea", "woman", "beach", "hat" ]
[ "cloud", "ship", "mountain", "tree", "lounge" ]
21
AMBER_21.jpg
Describe this image.
[ "ground", "dog" ]
[ "bench", "grass", "sky", "streetlamp", "person" ]
22
AMBER_22.jpg
Describe this image.
[ "mountain", "sky", "sea", "woman", "beach" ]
[ "ship", "umbrella", "car", "sun", "fish" ]
23
AMBER_23.jpg
Describe this image.
[ "sky", "cloud", "person", "desert", "crutch" ]
[ "sun", "bird", "car", "tree" ]
24
AMBER_24.jpg
Describe this image.
[ "house", "ball", "grass", "sign", "man", "tree", "fence", "road", "volleyball", "pulp" ]
[ "car", "woman", "dustbin", "bench", "sun" ]
25
AMBER_25.jpg
Describe this image.
[ "car", "river", "forest", "ground", "stone", "sky" ]
[ "sun", "bird", "flower", "person", "insect" ]
26
AMBER_26.jpg
Describe this image.
[ "forest", "child", "grass", "horse", "road", "deer" ]
[ "sun", "bird", "man", "flower", "insect" ]
27
AMBER_27.jpg
Describe this image.
[ "ground", "motorbike", "person", "hat", "flower", "grass", "road", "glove" ]
[ "sun", "sign", "tree", "bird", "house" ]
28
AMBER_28.jpg
Describe this image.
[ "ground", "person", "sky", "cloud", "forest", "road", "grass" ]
[ "sun", "bird", "sign", "car", "house" ]
29
AMBER_29.jpg
Describe this image.
[ "ground", "motorbike", "person", "hat", "forest", "glove", "road", "rail" ]
[ "sun", "sign", "tree", "bird", "house" ]
30
AMBER_30.jpg
Describe this image.
[ "sky", "cloud", "child", "grass", "flower" ]
[ "sun", "man", "bird", "tree", "car" ]
31
AMBER_31.jpg
Describe this image.
[ "beach", "child", "bird", "sea", "sky", "necklace" ]
[ "ship", "mountain", "tree", "lounge", "reef" ]
32
AMBER_32.jpg
Describe this image.
[ "sky", "sea", "beach", "dog" ]
[ "mountain", "ship", "person", "tree", "lounge" ]
33
AMBER_33.jpg
Describe this image.
[ "dog", "person", "bicycle", "ground", "road", "rope" ]
[ "sky", "tree", "grass", "streetlamp", "building" ]
34
AMBER_34.jpg
Describe this image.
[ "football", "person", "glass", "grass", "watch", "child" ]
[ "sun", "cloud", "sky", "bird", "tree" ]
35
AMBER_35.jpg
Describe this image.
[ "sky", "cloud", "person", "sea", "backpack", "ground", "guardrail", "sun", "railing" ]
[ "sun", "bench", "fish", "bird" ]
36
AMBER_36.jpg
Describe this image.
[ "person", "snow", "shoe", "hat" ]
[ "sky", "tree", "house", "streetlamp", "dog" ]
37
AMBER_37.jpg
Describe this image.
[ "person", "forest", "snow", "ground", "road" ]
[ "sign", "streetlamp", "bench", "bicycle", "dog" ]
38
AMBER_38.jpg
Describe this image.
[ "grass", "court", "person", "football" ]
[ "net", "sky", "cloud", "ground" ]
39
AMBER_39.jpg
Describe this image.
[ "sky", "cloud", "person", "bicycle", "sea", "beach" ]
[ "mountain", "ship", "lounge", "reef", "tree" ]
40
AMBER_40.jpg
Describe this image.
[ "person", "forest", "grass", "flower", "dog" ]
[ "sky", "cloud", "house", "ground", "sign" ]
41
AMBER_41.jpg
Describe this image.
[ "person", "wall", "ground" ]
[ "sky", "cloud", "grass", "sign", "dog" ]
42
AMBER_42.jpg
Describe this image.
[ "sky", "forest", "grass", "dog", "house", "stone" ]
[ "person", "cloud", "bench", "sign", "toy" ]
43
AMBER_43.jpg
Describe this image.
[ "sky", "cloud", "woman", "sheep", "grass", "tree", "ground", "road", "child" ]
[ "bird", "car", "sign", "dog", "motorbike" ]
44
AMBER_44.jpg
Describe this image.
[ "man", "grass", "road", "ground", "forest", "hat" ]
[ "sun", "sky", "dog", "sign", "cloud" ]
45
AMBER_45.jpg
Describe this image.
[ "dog", "toy", "frisbee", "person", "grass" ]
[ "sun", "sky", "cloud", "man", "ball" ]
46
AMBER_46.jpg
Describe this image.
[ "watermelon", "floor", "person", "chair", "child" ]
[ "wall", "window", "table", "ceiling", "door" ]
47
AMBER_47.jpg
Describe this image.
[ "sky", "sea", "person", "umbrella", "mountain", "beach" ]
[ "lounge", "ship", "surfboard", "dog", "tree" ]
48
AMBER_48.jpg
Describe this image.
[ "woman", "beach" ]
[ "sky", "sun", "sea", "dog", "lounge" ]
49
AMBER_49.jpg
Describe this image.
[ "sky", "sea", "person", "beach", "cloud", "hat", "child" ]
[ "lounge", "dog", "ship", "tree", "mountain" ]
50
AMBER_50.jpg
Describe this image.
[ "sky", "forest", "fence", "snow", "ground", "road", "man", "car", "hat", "glass", "glove", "mountain" ]
[ "cloud", "sign", "bus", "dog", "building" ]
51
AMBER_51.jpg
Describe this image.
[ "sky", "cloud", "person", "sea", "mountain", "beach" ]
[ "ship", "tree", "lounge", "umbrella", "dog" ]
52
AMBER_52.jpg
Describe this image.
[ "ball", "child" ]
[ "cup", "chair", "table", "door" ]
53
AMBER_53.jpg
Describe this image.
[ "dog", "flower", "grass", "necklace" ]
[ "sun", "sky", "cloud", "man", "woman" ]
54
AMBER_54.jpg
Describe this image.
[ "man", "guitar", "grass", "sky", "glass" ]
[ "sun", "cloud", "dog", "backpack" ]
55
AMBER_55.jpg
Describe this image.
[ "dog", "road", "grass" ]
[ "sun", "sky", "bird", "cloud", "person" ]
56
AMBER_56.jpg
Describe this image.
[ "grass", "ground", "road", "child", "bicycle" ]
[ "sign", "house", "bin", "sky", "cloud" ]
57
AMBER_57.jpg
Describe this image.
[ "sky", "forest", "grass", "ground", "road", "person" ]
[ "sign", "bench", "bin", "dog", "bicycle" ]
58
AMBER_58.jpg
Describe this image.
[ "cow", "person", "sea", "beach", "hat" ]
[ "sun", "cloud", "bird", "sky", "fish" ]
59
AMBER_59.jpg
Describe this image.
[ "woman", "hat", "flower", "flowerpot", "wall", "ground", "basket", "bucket" ]
[ "sun", "cloud", "bird", "dog", "building" ]
60
AMBER_60.jpg
Describe this image.
[ "sky", "forest", "grass", "dog", "ball", "football" ]
[ "cloud", "sun", "person", "tent", "lounge" ]
61
AMBER_61.jpg
Describe this image.
[ "person", "wall", "book", "notebook", "pen", "table", "ring", "earing", "book" ]
[ "phone", "cup", "light", "chair", "floor" ]
62
AMBER_62.jpg
Describe this image.
[ "dog", "building", "grass", "ground", "bench" ]
[ "sky", "cloud", "tree", "person", "house" ]
63
AMBER_63.jpg
Describe this image.
[ "sky", "mountain", "man", "sea" ]
[ "seegull", "ship", "beach", "tree", "lounge" ]
64
AMBER_64.jpg
Describe this image.
[ "sea", "beach", "dog", "necklace" ]
[ "ship", "sky", "mountain", "toy", "person" ]
65
AMBER_65.jpg
Describe this image.
[ "sky", "cloud", "person", "camera" ]
[ "grass", "flower", "tree", "dog", "house" ]
66
AMBER_66.jpg
Describe this image.
[ "forest", "grass", "dog", "ball", "tennis" ]
[ "sky", "sun", "person", "tent", "lounge" ]
67
AMBER_67.jpg
Describe this image.
[ "wall", "man", "ground", "cup" ]
[ "sky", "sun", "streetlamp", "building", "bench" ]
68
AMBER_68.jpg
Describe this image.
[ "mountain", "sky", "cloud", "ground", "horse", "person", "hat", "mountain", "grass" ]
[ "sun", "flower", "tree", "bird", "river" ]
69
AMBER_69.jpg
Describe this image.
[ "forest", "car", "ground", "person", "road", "water" ]
[ "sun", "sky", "cloud", "flower", "electrombile" ]
70
AMBER_70.jpg
Describe this image.
[ "sky", "cloud", "house", "fence", "tree", "woman", "bicycle", "ground", "road", "wall" ]
[ "sun", "grass", "dog", "sign" ]
71
AMBER_71.jpg
Describe this image.
[ "sky", "grass", "flower", "tree", "child" ]
[ "cloud", "bird", "tent", "dog", "toy" ]
72
AMBER_72.jpg
Describe this image.
[ "dog", "grass", "tree", "rope", "necklace" ]
[ "person", "sky", "cloud", "toy", "house" ]
73
AMBER_73.jpg
Describe this image.
[ "wall", "grass", "ground", "road", "sign", "man" ]
[ "sky", "sun", "tree", "car", "bin" ]
74
AMBER_74.jpg
Describe this image.
[ "sky", "cloud", "forest", "person", "hat", "glove", "motorbike" ]
[ "grass", "ground", "sign", "car", "house" ]
75
AMBER_75.jpg
Describe this image.
[ "ground", "road", "hydrant", "dog", "rope" ]
[ "sky", "cloud", "building", "streetlamp", "person" ]
76
AMBER_76.jpg
Describe this image.
[ "sky", "cloud", "mountain", "river", "person", "ship", "paddle", "hat" ]
[ "fish", "bird", "grass", "bridge" ]
77
AMBER_77.jpg
Describe this image.
[ "beach", "child", "toy", "shovel" ]
[ "sea", "sky", "cloud", "ship", "dog" ]
78
AMBER_78.jpg
Describe this image.
[ "sky", "plane", "person", "ground", "car", "mountain", "road", "stair" ]
[ "sun", "glass", "umbrella", "backpack" ]
79
AMBER_79.jpg
Describe this image.
[ "man", "football", "grass", "court" ]
[ "net", "sky", "sun", "tree", "cloud" ]
80
AMBER_80.jpg
Describe this image.
[ "sky", "forest", "grass", "child", "ground", "road" ]
[ "sign", "cloud", "dog", "house", "bird" ]
81
AMBER_81.jpg
Describe this image.
[ "sky", "grass", "football", "person" ]
[ "net", "cloud", "dog", "house", "tree" ]
82
AMBER_82.jpg
Describe this image.
[ "sea", "surfboard", "man" ]
[ "sky", "cloud", "mountain", "ship", "beach" ]
83
AMBER_83.jpg
Describe this image.
[ "sky", "sea", "beach", "person", "toy", "ball", "child" ]
[ "cloud", "mountain", "ship", "surfboard", "lounge" ]
84
AMBER_84.jpg
Describe this image.
[ "wall", "shampoo", "child", "bath", "toy", "cup", "water", "bottle" ]
[ "light", "floor", "toilet", "sink", "tap" ]
85
AMBER_85.jpg
Describe this image.
[ "wall", "bath", "tap", "man", "water" ]
[ "shampoo", "floor", "toilet", "light", "sink" ]
86
AMBER_86.jpg
Describe this image.
[ "dog", "ball", "grass" ]
[ "sun", "cloud", "flower", "person", "cup" ]
87
AMBER_87.jpg
Describe this image.
[ "sea", "beach", "person", "horse", "hat", "saddle" ]
[ "sky", "cloud", "mountain", "ship", "tree" ]
88
AMBER_88.jpg
Describe this image.
[ "hat", "sea", "ground", "rock", "bicycle", "telescope", "sky", "cloud", "person", "board" ]
[ "sun", "bird", "tree", "flower", "cup" ]
89
AMBER_89.jpg
Describe this image.
[ "sky", "cloud", "sea", "ship", "mountain", "beach", "man", "glass" ]
[ "sun", "tree", "lounge", "umbrella", "dog" ]
90
AMBER_90.jpg
Describe this image.
[ "dog", "bottle", "snow", "necklace" ]
[ "sky", "cloud", "person", "sign", "building" ]
91
AMBER_91.jpg
Describe this image.
[ "backpack", "grass", "person", "camera" ]
[ "sky", "cloud", "tree", "dog", "house" ]
92
AMBER_92.jpg
Describe this image.
[ "person", "hat", "motorbike", "stair", "ground", "glove", "road" ]
[ "sun", "sky", "cloud", "car", "sign" ]
93
AMBER_93.jpg
Describe this image.
[ "lake", "dog", "toy", "tennis", "ball" ]
[ "sky", "cloud", "mountain", "person" ]
94
AMBER_94.jpg
Describe this image.
[ "sky", "cloud", "mountain", "person", "surfboard", "sea" ]
[ "sun", "beach", "ship", "lounge", "dog" ]
95
AMBER_95.jpg
Describe this image.
[ "sky", "forest", "ground", "bus", "person", "grass", "road", "star" ]
[ "backpack", "house", "cloud", "dog", "car" ]
96
AMBER_96.jpg
Describe this image.
[ "sky", "cloud", "sun", "forest", "grass", "bicycle", "person", "stone", "hat", "mountain" ]
[ "bird", "dog", "camera", "sign" ]
97
AMBER_97.jpg
Describe this image.
[ "wall", "door", "man", "chair", "ground", "cone", "sign", "road" ]
[ "sky", "sun", "tree", "grass", "dog" ]
98
AMBER_98.jpg
Describe this image.
[ "beach", "man", "volleyball" ]
[ "net", "sea", "lounge", "umbrella", "tree" ]
99
AMBER_99.jpg
Describe this image.
[ "sky", "cloud", "grass", "ground", "road", "bicycle", "woman", "hat", "river" ]
[ "sun", "bird", "sign", "dog", "tree" ]
100
AMBER_100.jpg
Describe this image.
[ "sky", "cloud", "sea", "beach", "person", "child" ]
[ "ship", "mountain", "lounge", "dog", "umbrella" ]
End of preview. Expand in Data Studio

Hallucination Mitigation

One repo with the 3 benchmarks commonly used to evaluate hallucination mitigation in vision-language models: POPE, CHAIR and AMBER. Images are embedded, so nothing else needs downloading. Each benchmark, and each POPE split, loads by name.

Benchmark Task Config Rows Unique images
POPE Yes/no questions on object presence pope 27,000 1,489
CHAIR Captions scored against COCO objects chair 500 test, 300 dev 800
AMBER Captions scored against object lists (generative task) amber 1,004 1,004

Quick start

from datasets import load_dataset

repo = "aryadomain/hallucination-mitigation"

pope = load_dataset(repo, "pope", split="test")                         # all 27,000 questions
one = load_dataset(repo, "pope_coco_random", split="test")              # 1 split: MS-COCO random
one = load_dataset(repo, "pope_aokvqa_popular", split="test")           # 1 split: A-OKVQA popular
gqa = load_dataset(repo, "pope_gqa", split="test")                      # 1 dataset: all 3 GQA splits
adv = load_dataset(repo, "pope_adversarial", split="test")              # 1 split type across all 3 datasets
chair = load_dataset(repo, "chair", split="test")                       # 500 evaluation images
amber = load_dataset(repo, "amber", split="test")                       # 1,004 images

POPE configs

Each config has one test split. Configs share the same files, so there is no duplicated data.

Config Dataset Split type Rows
pope all 3 all 3 27,000
pope_coco, pope_aokvqa, pope_gqa 1 all 3 9,000 each
pope_random, pope_popular, pope_adversarial all 3 1 9,000 each
pope_coco_random MS-COCO random 3,000
pope_coco_popular MS-COCO popular 3,000
pope_coco_adversarial MS-COCO adversarial 3,000
pope_aokvqa_random A-OKVQA random 3,000
pope_aokvqa_popular A-OKVQA popular 3,000
pope_aokvqa_adversarial A-OKVQA adversarial 3,000
pope_gqa_random GQA random 3,000
pope_gqa_popular GQA popular 3,000
pope_gqa_adversarial GQA adversarial 3,000

Each of the 9 splits has 1,500 yes and 1,500 no questions. Question ids alternate between the 2 labels, so shuffle before taking a subset or sharding by stride; otherwise a subset can end up with a single label.

Fields

POPE

Field Type Meaning
question_id int Unique across the 27,000 questions, 1 to 27000
orig_id int question_id in the original split file
image, image_name image, string The image and its original file name
text string The question
label string yes or no
source string coco, aokvqa or gqa
category string random, popular or adversarial

CHAIR

Field Type Meaning
image_id int COCO image id
image, image_name image, string The image and its COCO file name
prompt string The prompt used in this project
objects list of strings COCO object categories annotated on the image
captions list of strings COCO reference captions

AMBER

Field Type Meaning
id int AMBER query id
image, image_name image, string The image and its AMBER file name
prompt string The AMBER query
truth list of strings Objects present, from AMBER's annotations
hallu list of strings Likely hallucinated objects, from AMBER's annotations

Protocol used in this project

Benchmark Prompt text Output
POPE <question> Please just answer yes or no. Greedy; first token decides
CHAIR Please describe this image in detail. Greedy; 64 and 512 tokens
AMBER Describe this image. Greedy

Wrap the text in your model's own chat template. On LLaVA-1.5 this is USER: <image>\n<text> ASSISTANT:.

Evaluation

POPE. evaluate_pope.py reports accuracy and F1 (yes is positive) for any config, per split, with a mean across splits when there is more than one. It expects 1 JSON object per line, such as {"question_id": 1, "answer": "yes"}, and stops if a question in the config has no prediction.

python evaluate_pope.py --config pope_coco_random --predictions predictions.jsonl
python evaluate_pope.py --config pope --predictions predictions.jsonl

CHAIR. scoring/chair/ holds the unchanged COCO instances_val2014.json and captions_val2014.json, plus chair_split.json with the test and dev ids. These are the inputs for a CHAIR scorer; this project used the chair.py distributed with OPERA. That scorer is not included here. The dev split is disjoint from test and is meant for method development; report results on test.

AMBER. scoring/amber/ holds the official inference.py and its data files, unchanged. See the AMBER repo for setup and usage. Only the generative task is included here as a config, as it is the one this project used.

Sources

  • POPE: the standard question files from the VCD repo, pope_* files for COCO, A-OKVQA and GQA. Note that some older POPE copies on the Hub differ from these on MS-COCO adversarial, A-OKVQA and GQA. Images come from COCO val2014 and GQA.
  • CHAIR: COCO val2014 images and annotations. The 800 ids were chosen for this project, so they may not match the 500 images used in other papers.
  • AMBER: the AMBER repo.

Image bytes and annotation contents are unchanged. Only the packaging differs.

Not included

The COCO train2014 subset used to rebuild a baseline method, AMBER's discriminative queries, and any model outputs or results.

Licences

This repo only packages data from other projects, and each part keeps its original terms. Check them before use, especially for commercial use.

Part Terms
COCO annotations CC BY 4.0, per the COCO terms
COCO images Flickr terms of use; COCO does not own the images
GQA images See the GQA site
POPE questions See the POPE and VCD repos
AMBER code and data files Apache 2.0 (scoring/amber/LICENSE)
AMBER images The AMBER repo does not state separate image terms

If you hold rights to any content here and want it removed, please open a discussion on this repo.

Citations

Please cite the original benchmarks:

  • POPE: Li et al., Evaluating Object Hallucination in Large Vision-Language Models, EMNLP 2023.
  • CHAIR: Rohrbach et al., Object Hallucination in Image Captioning, EMNLP 2018.
  • AMBER: Wang et al., An LLM-free Multi-dimensional Benchmark for MLLMs Hallucination Evaluation, arXiv:2311.07397, 2023.
  • COCO: Lin et al., Microsoft COCO: Common Objects in Context, ECCV 2014.
  • GQA: Hudson and Manning, GQA: Visual Reasoning and Compositional Question Answering, CVPR 2019.
  • A-OKVQA: Schwenk et al., A-OKVQA: A Benchmark for Visual Question Answering Using World Knowledge, ECCV 2022.
  • VCD: Leng et al., Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding, CVPR 2024.
Downloads last month
204

Paper for aryadomain/hallucination-mitigation