Tran1312/Image-Caption

BLIP3o Long-Caption 100K Image-Text Subset

This dataset is a locally reorganized subset of BLIP3o/BLIP3o-Pretrain-Long-Caption, containing approximately 100,000 image-text pairs selected from the original BLIP3o long-caption pretraining dataset [1].

The original BLIP3o long-caption collection contains approximately 27 million images, each paired with a long caption of roughly 120 tokens generated using Qwen2.5-VL-7B-Instruct [1].

The BLIP3-o project was introduced as part of a family of open unified multimodal models, together with associated training recipes, model weights, and datasets [2].

Important: This 100K dataset is a derived subset and local reorganization of BLIP3o/BLIP3o-Pretrain-Long-Caption. It is not an official standalone BLIP3o release. The directory layout, file naming scheme, and JSON metadata format described below are specific to this local dataset organization.

1. Dataset Source

The source dataset is:

BLIP3o/BLIP3o-Pretrain-Long-Caption

Official dataset repository:

https://huggingface.co/datasets/BLIP3o/BLIP3o-Pretrain-Long-Caption

According to the official dataset card, the source collection contains approximately 27 million image-caption pairs and is distributed primarily in WebDataset format [1].

This local dataset extracts only about 100,000 samples and converts them into a simpler filesystem layout consisting of:

1 image directory
+
1 JSON metadata file

The conversion is intended to simplify local experimentation and multimodal pretraining pipelines.

2. Directory Structure

The dataset is organized as:

dataset/
├── images/
│   ├── 000001.jpg
│   ├── 000002.jpg
│   ├── 000003.jpg
│   ├── ...
│   └── 100000.jpg
└── metadata.json

The exact root directory name is implementation-dependent.

All extracted images are stored directly inside the image directory rather than retaining the original WebDataset .tar shard structure.

The JSON metadata file stores the relationship between each image and its corresponding long-form caption.

3. Metadata Format

The metadata is stored as a JSON array.

Each element represents one image-text pair:

[
  {
    "dataset_name": "BLIP3o/BLIP3o-Pretrain-Long-Caption",
    "image_name": "000001.jpg",
    "text": "The image depicts a person in military attire, including a helmet and tactical gear, running through a grassy field dotted with yellow flowers. The individual appears to be in motion, possibly during a training exercise or a tactical operation. The background shows a mix of greenery and a dirt path, suggesting an open outdoor setting. The person's posture and the way they are holding their hands up might indicate a gesture of surrender or signaling. The overall scene conveys a sense of action and movement in a natural environment."
  },
  {
    "dataset_name": "BLIP3o/BLIP3o-Pretrain-Long-Caption",
    "image_name": "000002.jpg",
    "text": "The image captures a picturesque coastal scene with a clear blue sky and scattered clouds. In the foreground, a calm body of water reflects the serene atmosphere. A dock extends into the water, with a flagpole and a small structure at its end. Beyond the dock, a marina is visible, filled with various boats and yachts moored along the pier. The midground features a charming town with white buildings adorned with red roofs, and a prominent bell tower rises above the skyline. In the background, a range of rugged mountains provides a dramatic backdrop, adding depth and grandeur to the landscape. The overall composition exudes tranquility and natural beauty, typical of a Mediterranean coastal town."
  }
]

4. Metadata Fields

Each sample contains three fields.

dataset_name

"dataset_name": "BLIP3o/BLIP3o-Pretrain-Long-Caption"

Identifies the upstream dataset from which the sample was obtained.

This value is retained explicitly so that the original source remains traceable after local extraction or after multiple datasets are combined into a larger multimodal training corpus.

image_name

"image_name": "000001.jpg"

The filename of the locally stored image.

The corresponding physical image path is obtained by joining the image directory with this value:

image_path = os.path.join(images_dir, sample["image_name"])

For example:

images_dir:
dataset/images/

image_name:
000001.jpg

resolves to:

dataset/images/000001.jpg

text

"text": "The image depicts ..."

Contains the long-form natural-language description associated with the image.

In the original BLIP3o long-caption dataset, captions were generated using Qwen2.5-VL-7B-Instruct and are approximately 120 tokens long on average [1].

These captions can be used as supervision for vision-language pretraining tasks such as:

• —image-text alignment;
• —image-text contrastive learning;
• —image-conditioned text generation;
• —multimodal representation learning;
• —vision-language instruction or pretraining pipelines.

5. Relationship Between Metadata and Images

The relationship between metadata and physical files is intentionally simple:

metadata.json
    │
    ├── image_name: "000001.jpg"
    │
    ▼
images/000001.jpg

Each metadata entry points directly to one local image through image_name.

Conceptually:

for sample in metadata:
    image_path = os.path.join(images_dir, sample["image_name"])
    caption = sample["text"]

No additional shard lookup, archive extraction, or path mapping is required at training time.

6. Local Reorganization

The upstream BLIP3o long-caption dataset is distributed in WebDataset format, where image-text samples are stored inside .tar shards [1].

The original representation is conceptually similar to:

sa_000000.tar
sa_000001.tar
sa_000002.tar
...

This local 100K subset converts selected samples into:

WebDataset shards
        │
        ▼
selected ~100K samples
        │
        ├── image
        │      ▼
        │   images/*.jpg
        │
        └── caption/source metadata
               ▼
           metadata.json

The transformation changes only the local storage representation used by this dataset.

It does not imply that the reformatted subset is an independent official release from the BLIP3o authors.

7. Intended Use

This subset is designed primarily for research and engineering experiments involving multimodal representation learning and vision-language pretraining where working with the full approximately 27-million-sample collection is unnecessary or computationally impractical.

Possible use cases include:

• —validating a multimodal training pipeline;
• —prototyping vision-language model architectures;
• —image-text contrastive learning;
• —image-text matching;
• —image-grounded language generation;
• —ablation studies;
• —small-scale pretraining experiments;
• —debugging data loaders and distributed training systems.

Because this subset contains only a small fraction of the original BLIP3o long-caption dataset, results obtained from it should not be treated as equivalent to training on the complete source dataset.

8. Provenance

The provenance chain is:

BLIP3o project
      │
      ▼
BLIP3o/BLIP3o-Pretrain-Long-Caption
      │
      ├── ~27M image-text pairs
      │
      └── long captions generated with Qwen2.5-VL-7B-Instruct
      │
      ▼
sample selection
      │
      ▼
~100K selected image-text pairs
      │
      ▼
local reorganization
      │
      ├── images/*.jpg
      └── metadata.json

The dataset_name field is deliberately preserved in every record to maintain explicit provenance to the upstream dataset.

9. License and Attribution

The official Hugging Face repository for BLIP3o/BLIP3o-Pretrain-Long-Caption currently declares the dataset license as:

Apache License 2.0

[1]

Users of this derived subset should preserve attribution to the original BLIP3o project and should review the license and dataset card of the upstream repository before redistribution or deployment.

This local subset does not claim ownership of the original BLIP3o data.

The local reformatting of the data into individual image files and JSON metadata does not transfer or replace any rights associated with the source material.

In particular, although the upstream Hugging Face repository declares an Apache-2.0 license, users redistributing image data should independently verify whether additional rights, terms, or restrictions apply to underlying third-party image content. The source dataset card should be treated as the authoritative reference for the upstream distribution.

Official source:

https://huggingface.co/datasets/BLIP3o/BLIP3o-Pretrain-Long-Caption

Apache License 2.0:

https://www.apache.org/licenses/LICENSE-2.0

10. Citation

If this subset is used in research, the original BLIP3-o work should be cited.

@article{chen2025blip3o,
  title   = {BLIP3-o: A Family of Fully Open Unified Multimodal Models--Architecture, Training and Dataset},
  author  = {Jiuhai Chen and Zhiyang Xu and Xichen Pan and Yushi Hu and
             Can Qin and Tom Goldstein and Lifu Huang and Tianyi Zhou and
             Saining Xie and Silvio Savarese and Le Xue and Caiming Xiong and
             Ran Xu},
  journal = {arXiv preprint arXiv:2505.09568},
  year    = {2025}
}

When reporting experiments using this specific 100K subset, the experimental setup should additionally state that only approximately 100,000 samples were extracted from BLIP3o/BLIP3o-Pretrain-Long-Caption and that the samples were locally reorganized into individual image files and JSON metadata.

For example:

We use a 100K-sample subset derived from the
BLIP3o/BLIP3o-Pretrain-Long-Caption dataset [1,2].
The selected samples are locally reorganized into individual JPEG files
and a JSON metadata file containing image-caption pairs.

11. References

[1] BLIP3o Team, BLIP3o-Pretrain-Long-Caption Dataset. Hugging Face Datasets. https://huggingface.co/datasets/BLIP3o/BLIP3o-Pretrain-Long-Caption

[2] J. Chen, Z. Xu, X. Pan, Y. Hu, C. Qin, T. Goldstein, L. Huang, T. Zhou, S. Xie, S. Savarese, L. Xue, C. Xiong, and R. Xu, “BLIP3-o: A Family of Fully Open Unified Multimodal Models—Architecture, Training and Dataset.” arXiv:2505.09568, 2025. https://arxiv.org/abs/2505.09568

[3] Apache Software Foundation, Apache License, Version 2.0. https://www.apache.org/licenses/LICENSE-2.0