Datasets:
CaptionExam v0.1: 109 images, 3,099 questions
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- LICENSE +13 -0
- README.md +101 -0
- images/img_0001.jpg +3 -0
- images/img_0002.jpg +3 -0
- images/img_0003.jpg +3 -0
- images/img_0004.jpg +3 -0
- images/img_0005.jpg +3 -0
- images/img_0006.jpg +3 -0
- images/img_0007.jpg +3 -0
- images/img_0008.jpg +3 -0
- images/img_0009.jpg +3 -0
- images/img_0010.jpg +3 -0
- images/img_0011.jpg +3 -0
- images/img_0012.jpg +3 -0
- images/img_0013.jpg +3 -0
- images/img_0014.jpg +3 -0
- images/img_0015.jpg +3 -0
- images/img_0016.jpg +3 -0
- images/img_0017.jpg +3 -0
- images/img_0018.jpg +3 -0
- images/img_0019.jpg +3 -0
- images/img_0020.jpg +3 -0
- images/img_0021.jpg +3 -0
- images/img_0022.jpg +3 -0
- images/img_0023.jpg +3 -0
- images/img_0024.jpg +3 -0
- images/img_0025.jpg +3 -0
- images/img_0026.jpg +3 -0
- images/img_0027.jpg +3 -0
- images/img_0028.jpg +3 -0
- images/img_0029.jpg +3 -0
- images/img_0030.jpg +3 -0
- images/img_0031.jpg +3 -0
- images/img_0032.jpg +3 -0
- images/img_0033.jpg +3 -0
- images/img_0034.jpg +3 -0
- images/img_0035.jpg +3 -0
- images/img_0036.jpg +3 -0
- images/img_0037.jpg +3 -0
- images/img_0038.jpg +3 -0
- images/img_0039.jpg +3 -0
- images/img_0040.jpg +3 -0
- images/img_0041.jpg +3 -0
- images/img_0042.jpg +3 -0
- images/img_0043.jpg +3 -0
- images/img_0044.jpg +3 -0
- images/img_0045.jpg +3 -0
- images/img_0046.jpg +3 -0
- images/img_0047.jpg +3 -0
- images/img_0048.jpg +3 -0
LICENSE
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
CaptionExam dataset terms
|
| 2 |
+
|
| 3 |
+
The questions, options, answers and region annotations in this dataset are
|
| 4 |
+
released for research use.
|
| 5 |
+
|
| 6 |
+
The images were collected from the public web. Their copyright stays with
|
| 7 |
+
their owners. They are provided only for non-commercial research and
|
| 8 |
+
evaluation, and may not be redistributed for any other purpose.
|
| 9 |
+
|
| 10 |
+
If you own an image in this dataset and want it removed, open an issue at
|
| 11 |
+
https://github.com/Ki-Seki/CaptionExam/issues and it will be taken down.
|
| 12 |
+
|
| 13 |
+
THE DATASET IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.
|
README.md
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
pretty_name: CaptionExam
|
| 3 |
+
license: other
|
| 4 |
+
license_name: research-only
|
| 5 |
+
license_link: LICENSE
|
| 6 |
+
language:
|
| 7 |
+
- en
|
| 8 |
+
task_categories:
|
| 9 |
+
- image-to-text
|
| 10 |
+
- visual-question-answering
|
| 11 |
+
tags:
|
| 12 |
+
- image-captioning
|
| 13 |
+
- dense-captioning
|
| 14 |
+
- hallucination
|
| 15 |
+
- benchmark
|
| 16 |
+
size_categories:
|
| 17 |
+
- 1K<n<10K
|
| 18 |
+
configs:
|
| 19 |
+
- config_name: questions
|
| 20 |
+
default: true
|
| 21 |
+
data_files:
|
| 22 |
+
- split: test
|
| 23 |
+
path: questions.jsonl
|
| 24 |
+
- config_name: images
|
| 25 |
+
data_files:
|
| 26 |
+
- split: test
|
| 27 |
+
path: images/*
|
| 28 |
+
---
|
| 29 |
+
|
| 30 |
+
# CaptionExam
|
| 31 |
+
|
| 32 |
+
A hard, fine-grained benchmark for image captions. A text-only judge reads a model's caption, never the image, and answers multiple-choice questions about the image. Every question also offers *"not mentioned / indeterminable"*, so each verdict says whether the caption got a fact right, left it out (**omission**), or got it wrong (**hallucination**).
|
| 33 |
+
|
| 34 |
+
Code, judge setup and leaderboard: [github.com/Ki-Seki/CaptionExam](https://github.com/Ki-Seki/CaptionExam)
|
| 35 |
+
|
| 36 |
+
| | |
|
| 37 |
+
|---|---|
|
| 38 |
+
| Images | 109: photos, illustrations, product shots, posters, book covers, text-heavy graphics |
|
| 39 |
+
| Questions | 3,099 multiple-choice, 15–42 per image (median 27) |
|
| 40 |
+
| Options | 3–6 per question, one of them `not mentioned / indeterminable` |
|
| 41 |
+
| Whole-image questions | 120 (`bbox` = `[0, 0, 1000, 1000]`); the rest target a region |
|
| 42 |
+
|
| 43 |
+
The questions cover objects and counts, colors and materials, pose and orientation, layout and spatial relations, rendered text and its exact wording, typography, graphic style, lighting, and background.
|
| 44 |
+
|
| 45 |
+
## Files
|
| 46 |
+
|
| 47 |
+
```text
|
| 48 |
+
questions.jsonl one question per line
|
| 49 |
+
images/<image_id>.jpg the images
|
| 50 |
+
images/metadata.jsonl image_id, width, height, n_questions
|
| 51 |
+
```
|
| 52 |
+
|
| 53 |
+
Each question:
|
| 54 |
+
|
| 55 |
+
```json
|
| 56 |
+
{"id": "img_0001_q02", "image_id": "img_0001",
|
| 57 |
+
"question": "What color is the body of the pickup truck parked under the carport?",
|
| 58 |
+
"options": ["white or light silver", "dark blue or black", "red or maroon", "dark green",
|
| 59 |
+
"not mentioned / indeterminable"],
|
| 60 |
+
"answer": ["white or light silver"],
|
| 61 |
+
"bbox": [453, 552, 622, 804]}
|
| 62 |
+
```
|
| 63 |
+
|
| 64 |
+
`bbox` is the region the question is about, as `[y_min, x_min, y_max, x_max]` normalized to 0–1000. The questions never mention a caption, so they can also grade a generated image, or a model answering while it looks at the original.
|
| 65 |
+
|
| 66 |
+
## Usage
|
| 67 |
+
|
| 68 |
+
```bash
|
| 69 |
+
pip install git+https://github.com/Ki-Seki/CaptionExam
|
| 70 |
+
captionexam download
|
| 71 |
+
captionexam caption --model <vlm> --name <name> --base-url <openai-compatible endpoint>
|
| 72 |
+
captionexam judge --judge-base-url <phi-4 endpoint>
|
| 73 |
+
captionexam board
|
| 74 |
+
```
|
| 75 |
+
|
| 76 |
+
Or read the files directly:
|
| 77 |
+
|
| 78 |
+
```python
|
| 79 |
+
from datasets import load_dataset
|
| 80 |
+
|
| 81 |
+
questions = load_dataset("Ki-Seki/CaptionExam", "questions", split="test")
|
| 82 |
+
images = load_dataset("Ki-Seki/CaptionExam", "images", split="test")
|
| 83 |
+
```
|
| 84 |
+
|
| 85 |
+
## How the questions were made
|
| 86 |
+
|
| 87 |
+
The questions were mined from the mistakes of real captioning models. A strong vision-language model looked at an image and one model's caption, then wrote questions about the facts that caption got wrong or left out. The wrong options had to be plausible for that particular image, so nobody can pick the answer without information about the image. Two other strong vision-language models then answered every candidate question while looking at the image. A question was kept only if both answered it correctly, and every kept question was reviewed by a person.
|
| 88 |
+
|
| 89 |
+
## Scoring
|
| 90 |
+
|
| 91 |
+
With the default judge ([Phi-4](https://huggingface.co/microsoft/phi-4), temperature 0, answers constrained to the option text), a caption's **accuracy** is the share of questions answered correctly. The **omission rate** is the share where the judge chose *not mentioned*, and the **hallucination rate** is the share where it chose a wrong option. Compare models only under the same judge and caption prompt.
|
| 92 |
+
|
| 93 |
+
## Limitations
|
| 94 |
+
|
| 95 |
+
- The judge makes mistakes. Model rankings agree closely with a strong proprietary judge (Spearman ρ ≈ 0.89 over 11 captioners), but a single verdict is not ground truth.
|
| 96 |
+
- Questions mined from captioners' errors favor the details those captioners missed. They probe how complete and faithful a caption is, not how fluent or well organized it is.
|
| 97 |
+
- 109 images is a small set; differences of a point or two between models are within noise.
|
| 98 |
+
|
| 99 |
+
## License and takedown
|
| 100 |
+
|
| 101 |
+
The questions and annotations are released for research use. The images were collected from the public web; their copyright stays with their owners. They are provided only for non-commercial research and evaluation. If you own an image and want it removed, open an issue at [github.com/Ki-Seki/CaptionExam](https://github.com/Ki-Seki/CaptionExam/issues) and we will take it down.
|
images/img_0001.jpg
ADDED
|
Git LFS Details
|
images/img_0002.jpg
ADDED
|
Git LFS Details
|
images/img_0003.jpg
ADDED
|
Git LFS Details
|
images/img_0004.jpg
ADDED
|
Git LFS Details
|
images/img_0005.jpg
ADDED
|
Git LFS Details
|
images/img_0006.jpg
ADDED
|
Git LFS Details
|
images/img_0007.jpg
ADDED
|
Git LFS Details
|
images/img_0008.jpg
ADDED
|
Git LFS Details
|
images/img_0009.jpg
ADDED
|
Git LFS Details
|
images/img_0010.jpg
ADDED
|
Git LFS Details
|
images/img_0011.jpg
ADDED
|
Git LFS Details
|
images/img_0012.jpg
ADDED
|
Git LFS Details
|
images/img_0013.jpg
ADDED
|
Git LFS Details
|
images/img_0014.jpg
ADDED
|
Git LFS Details
|
images/img_0015.jpg
ADDED
|
Git LFS Details
|
images/img_0016.jpg
ADDED
|
Git LFS Details
|
images/img_0017.jpg
ADDED
|
Git LFS Details
|
images/img_0018.jpg
ADDED
|
Git LFS Details
|
images/img_0019.jpg
ADDED
|
Git LFS Details
|
images/img_0020.jpg
ADDED
|
Git LFS Details
|
images/img_0021.jpg
ADDED
|
Git LFS Details
|
images/img_0022.jpg
ADDED
|
Git LFS Details
|
images/img_0023.jpg
ADDED
|
Git LFS Details
|
images/img_0024.jpg
ADDED
|
Git LFS Details
|
images/img_0025.jpg
ADDED
|
Git LFS Details
|
images/img_0026.jpg
ADDED
|
Git LFS Details
|
images/img_0027.jpg
ADDED
|
Git LFS Details
|
images/img_0028.jpg
ADDED
|
Git LFS Details
|
images/img_0029.jpg
ADDED
|
Git LFS Details
|
images/img_0030.jpg
ADDED
|
Git LFS Details
|
images/img_0031.jpg
ADDED
|
Git LFS Details
|
images/img_0032.jpg
ADDED
|
Git LFS Details
|
images/img_0033.jpg
ADDED
|
Git LFS Details
|
images/img_0034.jpg
ADDED
|
Git LFS Details
|
images/img_0035.jpg
ADDED
|
Git LFS Details
|
images/img_0036.jpg
ADDED
|
Git LFS Details
|
images/img_0037.jpg
ADDED
|
Git LFS Details
|
images/img_0038.jpg
ADDED
|
Git LFS Details
|
images/img_0039.jpg
ADDED
|
Git LFS Details
|
images/img_0040.jpg
ADDED
|
Git LFS Details
|
images/img_0041.jpg
ADDED
|
Git LFS Details
|
images/img_0042.jpg
ADDED
|
Git LFS Details
|
images/img_0043.jpg
ADDED
|
Git LFS Details
|
images/img_0044.jpg
ADDED
|
Git LFS Details
|
images/img_0045.jpg
ADDED
|
Git LFS Details
|
images/img_0046.jpg
ADDED
|
Git LFS Details
|
images/img_0047.jpg
ADDED
|
Git LFS Details
|
images/img_0048.jpg
ADDED
|
Git LFS Details
|