Datasets:

deadbear34
/

Exodus_v1

license: mit
task_categories:
  - image-to-text
language:
  - en
tags:
  - dense-captioning
  - joycaption
  - laion
  - archenix
  - exodus
size_categories:
  - 100K<n<1M

Exodus v1

504,504 image-text pairs combining original LAION images with detailed ~260-word descriptions generated by JoyCaption Beta One.

Dataset Statistics

Metric	Value
Total images	504,504
Dense captions	504,504
Coverage	100.0%
Avg dense caption	223 words
Median dense caption	223 words
Dense range	127-360 words
IQR (25th-75th)	216-230 words
Avg original caption	11 words

Word Count Distribution

Caption Length Buckets

Range	Count	Percentage
<50	0	0.0%
50-100	0	0.0%
100-150	14	0.0%
150-200	3,736	0.7%
200-250	497,393	98.6%
250-300	3,293	0.7%
300+	68	0.0%

Word Cloud Comparison

Left: Dense captions (rich, descriptive vocabulary). Right: Original LAION captions (short, keyword-style).

Columns

Column	Type	Description
`image`	PIL Image	Original image from LAION
`original_caption`	string	Short original LAION caption
`dense_caption`	string	Detailed ~260 word JoyCaption description

Usage

from datasets import load_dataset

ds = load_dataset("deadbear34/Exodus_v1", split="train")

# Access a sample
print(ds[0]["dense_caption"])
ds[0]["image"].show()

# Filter by caption length
long_captions = ds.filter(lambda x: len(x["dense_caption"].split()) > 200)

Generation Details

Parameter	Value
Source dataset	deadbear34/laion-exodus-500k
Captioning model	JoyCaption Beta One
Caption type	Descriptive
Max words	260
Temperature	0.6
Top-p	0.9
Max tokens	512
Hardware	NVIDIA A100-80GB
Generation time	~16 hours
Speed	~8.8 images/second

Citation

If you use this dataset, please cite:

@misc{exodus_v1_2026,
  title={Exodus v1: LAION Images with Dense JoyCaption Descriptions},
  author={Archenix Team},
  year={2026},
  url={https://huggingface.co/datasets/deadbear34/Exodus_v1}
}

License

MIT License. Images are sourced from LAION and may be subject to their original copyright.