prithivMLmods/OpenCaption-Unified-10K

OpenCaption-Unified-10K

OpenCaption-Unified-10K is a dense image captioning dataset built from 10,000 images paired with long-form synthetic captions generated using the Qwen3.5 multimodal model. Each caption is produced through a dedicated Qwen3.5 captioning pipeline designed to generate detailed, high-fidelity descriptions of scene composition, subject attributes, spatial relationships, activities, and overall visual context rather than short, generic captions. The dataset is intended for training and evaluating image-to-text, vision-language, and dense captioning models that require rich textual grounding. The dataset is built primarily from publicly available images, which make up the majority of the input imagery, together with additional publicly available datasets. Every image is paired with a single comprehensive caption that attempts to describe the complete visual scene, making the dataset suitable for supervised vision-language training, instruction tuning, caption refinement, retrieval, and multimodal research.

Dataset Statistics

Dataset Structure

Each sample contains the following fields:

Example:

sample = ds[0]

print(sample.keys())

# dict_keys([
#     "image",
#     "response"
# ])

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset(
    "prithivMLmods/OpenCaption-Unified-10K",
    split="train"
)

Example Usage

from datasets import load_dataset
import matplotlib.pyplot as plt

dataset = load_dataset(
    "prithivMLmods/OpenCaption-Unified-10K",
    split="train"
)

sample = dataset[0]

image = sample["image"]
caption = sample["response"]

print(caption)

plt.figure(figsize=(8, 8))
plt.imshow(image)
plt.axis("off")
plt.show()

Citation

@misc{prithiv_sakthi_2026,
	author       = { Prithiv Sakthi },
	title        = { OpenCaption-Unified-10K (Revision 75cba9f) },
	year         = 2026,
	url          = { https://huggingface.co/datasets/prithivMLmods/OpenCaption-Unified-10K },
	doi          = { 10.57967/hf/9578 },
	publisher    = { Hugging Face }
}

License

This dataset is released under the Apache-2.0 License.