| --- |
| license: apache-2.0 |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train-* |
| dataset_info: |
| features: |
| - name: image |
| list: image |
| - name: response |
| dtype: string |
| splits: |
| - name: train |
| num_bytes: 4444062367 |
| num_examples: 10602 |
| download_size: 4431804584 |
| dataset_size: 4444062367 |
| task_categories: |
| - image-text-to-text |
| - image-to-text |
| language: |
| - en |
| pretty_name: MultiImage |
| size_categories: |
| - 10K<n<100K |
| tags: |
| - image-captioning |
| - dense-captioning |
| - computer-vision |
| - multimodal |
| - synthetic-captions |
| - deep-caption |
| - image |
| - vlm |
| --- |
| |
| ## **MultiImage-Caption** |
|
|
| **MultiImage-Caption** is a multimodal dense captioning dataset featuring **10,602** entries designed for training, supervised fine-tuning (SFT), and evaluating multi-image Vision-Language Models (such as Qwen2-VL, LLaVA-NeXT, and PaliGemma). |
|
|
| Each sample pairs an interleaved set of multiple images (`images list`) with comprehensive, comparative, and granular descriptive captions (`response`) analyzing the context, subjects, and interactions across all provided images. |
|
|
| - **Curator:** [prithivMLmods](https://huggingface.co/prithivMLmods) |
| - **Total Samples:** 10,602 rows |
| - **Total Size:** ~4.43 GB |
| - **Format:** Parquet (`image`, `response`) |
| - **Modalities:** Image, Text |
| - **Split:** Train |
|
|
| ## Dataset Structure & Schema |
|
|
| ### Feature Fields |
|
|
| | Field | Type | Description | |
| | :--- | :--- | :--- | |
| | `image` | `Sequence[Image]` | A list of multiple target RGB images grouped together per entry | |
| | `response` | `string` | Dense, structured textual explanation breaking down and comparing each individual image | |
|
|
| ### Example Response Structure |
|
|
| ```text |
| Here is a detailed explanation of each image: |
| |
| **Image 1**: This photograph captures an energetic moment at an outdoor event... |
| **Image 2**: A contrasting perspective showing subjects interacting with their surroundings... |
| ``` |
|
|
| ## How to Use |
|
|
| ### Loading with `datasets` |
|
|
| ```python |
| from datasets import load_dataset |
| |
| # Load dataset from the Hugging Face Hub |
| dataset = load_dataset("prithivMLmods/MultiImage-Caption", split="train") |
| |
| # Access a single sample |
| sample = dataset[0] |
| images = sample["image"] # List of PIL Images |
| response = sample["response"] # Multi-image detailed caption |
| |
| print(f"Number of images in sample: {len(images)}") |
| print("Response preview:\n", response[:250]) |
| ``` |
|
|
| ### Formatting for Multi-Image VLM SFT |
|
|
| ```python |
| def format_multi_image_conversation(example): |
| num_images = len(example["image"]) |
| image_tokens = "".join([f"<image_{i+1}>\n" for i in range(num_images)]) |
| |
| prompt = ( |
| f"{image_tokens}Provide a detailed, step-by-step description and comparative " |
| "analysis of each of the provided images." |
| ) |
| |
| return { |
| "images": example["image"], |
| "prompt": prompt, |
| "completion": example["response"] |
| } |
| |
| ``` |
|
|
| ## Intended Uses |
|
|
| * **Multi-Image Reasoning:** Training models to correlate, compare, and reason over sequences of visual inputs simultaneously. |
| * **Dense Captioning:** Generating rich, descriptive long-form visual commentary instead of brief single-sentence captions. |
| * **Interleaved Multimodal Instruction-Tuning:** Building datasets for conversational agents handling multi-image document analysis, video keyframes, or side-by-side visual comparisons. |
|
|
| ## License |
|
|
| This dataset is distributed under the **Apache-2.0 License**. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{prithivmlmods2026multiimagecaption, |
| title = {MultiImage-Caption: A Dense Multi-Image Multimodal Dataset}, |
| author = {prithivMLmods}, |
| year = {2026}, |
| publisher = {Hugging Face}, |
| howpublished = {\url{https://huggingface.co/datasets/prithivMLmods/MultiImage-Caption}} |
| } |
| ``` |
