prithivMLmods's picture
Update README.md
1d572e3 verified
|
Raw
History Blame Contribute Delete
7.32 kB
metadata
dataset_info:
  features:
    - name: image
      dtype: image
    - name: image_source
      dtype: string
    - name: response
      dtype: string
  splits:
    - name: train
      num_bytes: 4004176789
      num_examples: 158958
  download_size: 3837976031
  dataset_size: 4004176789
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
tags:
  - agent-traces
  - image-captioning
  - dense-captioning
  - uhd
  - computer-vision
  - multimodal
  - synthetic-captions
  - deep-caption
  - image
license: apache-2.0
task_categories:
  - image-text-to-text
  - image-to-text
language:
  - en
pretty_name: MultiCaptions
size_categories:
  - 100K<n<1M

OpenCOCO-I2T-Repack-Tiny

OpenCOCO-I2T-Repack-Tiny is a compact image-to-text / image-text-to-text captioning dataset containing 158,958 image samples sourced from the COCO dataset and repackaged into a lightweight format suitable for vision-language model (VLM) fine-tuning. The dataset contains synthesized responses generated using a custom Qwen3.5 multimodal captioning pipeline. The input images undergo lossless image compression to significantly reduce the overall storage footprint while preserving the visual information required for training and inference. With approximately 159K samples compressed into only 3.84 GB, OpenCOCO-I2T-Repack-Tiny is designed to provide an easily accessible and efficient dataset for VLM-based fine-tuning, experimentation, and image understanding tasks.

Dataset Overview

Property Details
Dataset Name OpenCOCO-I2T-Repack-Tiny
Samples 158,958
Task Image-to-Text / Image-Text-to-Text
Modality Image + Text
Image Source COCO
Caption Generation Qwen3.5 Custom Multimodal Pipeline
Format Parquet
Dataset Size 3.84 GB
License Apache-2.0
Primary Use VLM Fine-Tuning / Image Understanding

Description

OpenCOCO-I2T-Repack-Tiny is a repackaged version of COCO-based image data designed specifically for efficient multimodal training workflows.

The dataset consists of:

  • 158,958 image-text samples
  • Images sourced from the publicly available COCO dataset
  • Rich textual responses generated through a custom Qwen3.5 multimodal pipeline
  • Losslessly compressed input images
  • A compact Parquet-based dataset structure
  • Image source metadata preserved alongside the generated responses

The goal is to provide a relatively small storage footprint while retaining a large number of image-text training examples.

Dataset Construction

The dataset construction pipeline consists of several stages:

  1. Image Collection

    • Images are sourced from the COCO dataset.
    • The original image source is retained through the image_source field.
  2. Multimodal Caption Generation

    • Images are processed through a custom Qwen3.5-based multimodal pipeline.
    • The pipeline generates descriptive textual responses based on the visual content of each image.
  3. Image Compression

    • Input images undergo lossless image compression.
    • The objective is to reduce storage requirements without introducing lossy visual degradation.
  4. Dataset Repacking

    • Images and generated responses are consolidated into a compact dataset structure.
    • The final dataset is packaged in Parquet format for efficient loading and processing.
  5. VLM Training Preparation

    • The resulting image-text pairs can be directly adapted for multimodal fine-tuning workflows.

Dataset Structure

The dataset contains the following primary fields:

image
image_source
response

image

The compressed input image used for multimodal captioning and analysis.

image_source

The source associated with the original image.

For COCO samples, the source is:

COCO Dataset

response

The synthesized textual response generated by the custom Qwen3.5 multimodal pipeline.

The responses are intended to provide rich visual descriptions suitable for image understanding and VLM training.

Example

image:
[COCO image]

image_source:
http://cocodataset.org/

response:
This outdoor wooden table scene features a hearty breakfast setup, bathed in bright sunlight. In the foreground, multiple plates and food items are arranged across the table...

Dataset Statistics

  • Total Samples: 158,958
  • Approximate Dataset Size: 3.84 GB
  • Source: COCO
  • Task: Image-to-Text / Image-Text-to-Text
  • Image Compression: Lossless
  • Caption Generation: Qwen3.5 Custom Multimodal Pipeline

The compact storage footprint makes the dataset convenient for local experimentation, cloud training, and VLM fine-tuning environments with limited storage capacity.

Intended Use

OpenCOCO-I2T-Repack-Tiny can be used for:

  • Vision-Language Model fine-tuning
  • Image captioning
  • Image understanding
  • Visual instruction tuning
  • Multimodal representation learning
  • Image-to-text generation
  • Image-text alignment
  • VLM benchmarking and experimentation
  • Lightweight multimodal training pipelines

Why OpenCOCO-I2T-Repack-Tiny?

The primary goal of this repack is efficiency.

Rather than maintaining a large raw image collection, the dataset combines a high sample count with aggressive lossless compression and an efficient Parquet-based representation.

This results in:

158,958 image-text samples in approximately 3.84 GB.

This makes the dataset particularly useful for researchers and developers who want a large number of multimodal examples without requiring an excessively large amount of local or cloud storage.

Data Source

The underlying images are sourced from the COCO dataset.

COCO Dataset

This dataset is a repackaged and processed resource containing generated textual responses and compressed image representations.

Users should review the original source dataset's terms and attribution requirements when using the data.

Citation

If you use OpenCOCO-I2T-Repack-Tiny in your research or project, please cite:

@misc{prithiv_sakthi_2026,
    author       = {Prithiv Sakthi},
    title        = {OpenCOCO-I2T-Repack-Tiny (Revision a7ae40c)},
    year         = {2026},
    url          = {https://huggingface.co/datasets/prithivMLmods/OpenCOCO-I2T-Repack-Tiny},
    doi          = {10.57967/hf/10057},
    publisher    = {Hugging Face}
}

Dataset page:

OpenCOCO-I2T-Repack-Tiny on Hugging Face

Dataset creator:

prithivMLmods on Hugging Face

License

This dataset is released under the Apache-2.0 license as indicated by the dataset repository.

Because the underlying imagery originates from COCO, users should also review the applicable terms, attribution requirements, and usage conditions associated with the original source material.