README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet

New Conversation

A large-scale multimodal dataset for image understanding, visual question answering, and image-to-text instruction tuning.

CrowdMind/new-conversation contains approximately 558K image-and-text training examples. Each example pairs an image with a natural-language instruction and an assistant response describing or interpreting the visual content.

The dataset focuses on teaching vision-language models to understand images and produce useful textual descriptions.

Dataset Overview

Property	Value
Dataset	CrowdMind/new-conversation
Examples	558,000
Splits	train
Modalities	Image + Text
Format	Parquet
Conversation structure	2 messages
Primary language	English
Primary task	Image-to-text understanding

What Is Included?

Each training example contains an image, a user instruction, and an assistant response.

The instructions use different ways of asking a model to interpret an image. Examples include:

• Describe the image.
• Summarize the visual content.
• Identify objects and subjects.
• Explain what is happening in a photograph.
• Describe important visual features.
• Provide a concise interpretation of an image.
• Identify products, artwork, locations, and other objects.

The dataset covers a broad range of visual subjects.

Categories of Visual Content

The following categories describe the types of content represented in the dataset. They are descriptive groupings, not pre-existing dataset labels.

Category	Examples
People & Activities	People, groups, clothing, sports, social activities
Animals & Wildlife	Birds, pets, insects, wildlife
Nature & Landscapes	Mountains, beaches, rivers, plants, scenery
Vehicles & Transportation	Cars, trucks, motorcycles, transportation
Buildings & Architecture	Houses, hotels, structures, interiors
Products & Objects	Consumer products, furniture, tools, accessories
Food & Dining	Meals, desserts, ingredients, food presentation
Art & Illustrations	Sculptures, drawings, paintings, illustrations
Signs & Text	Signs, logos, lettering, advertisements
Fashion & Apparel	Clothing, shoes, jewelry, accessories
Technology	Electronics, equipment, devices
Events & Entertainment	Concerts, celebrations, performances

Dataset Structure

The Hugging Face viewer exposes three main fields:

• image: The image associated with the example.
• messages: A structured conversation using user and assistant roles.

Training Applications

This dataset can be used in several stages of multimodal model development.

1. Vision-Language SFT

Use the image and associated conversation as supervised fine-tuning examples.

The model learns to associate visual information with natural-language instructions and responses.

2. Image Captioning

Train models to generate concise descriptions of images.

Potential capabilities include:

• Object identification.
• Scene descriptions.
• Attribute recognition.
• Visual detail extraction.
• Natural-language image summaries.

3. Visual Question Answering

The question-and-answer structure provides a starting point for training models to respond to image-related questions.

Additional question generation could expand the dataset beyond general image descriptions.

4. Multimodal Pretraining and Instruction Tuning

The dataset can supplement larger multimodal training collections, particularly during instruction tuning.

It can help develop the connection between visual encoders and language-model representations.

Data Quality Considerations

The dataset contains a large variety of image-description examples, but the visible samples also show some quality variation.

Potential issues include:

• Incomplete or truncated descriptions.
• Incorrect object identification.
• Repetitive instruction templates.
• Keyword-heavy responses rather than natural sentences.
• Occasional grammatical errors.
• Descriptions that do not fully explain the image.

For best results, consider implementing:

1. Duplicate detection.
2. Image-caption consistency checks.
3. Text quality filtering.
4. Language normalization.
5. Caption length filtering.
6. Automated and human-assisted quality evaluation.

Do not assume that every response is a verified description of its associated image.

Limitations

• The published dataset currently exposes a single training split.
• The examples use predominantly English instructions.
• Most visible conversations contain only one user message and one assistant response.
• The dataset should not be treated as a general-purpose multi-turn conversational dataset.
• The quality and accuracy of individual image descriptions can vary.
• Additional validation and test splits should be created for controlled experiments.

Intended Use

This dataset is intended for researchers and developers working on:

• Vision-language models (VLMs).
• Multimodal large language models.
• Image captioning systems.
• Visual instruction following.
• Image understanding.
• Multimodal supervised fine-tuning.

Citation

If you use this dataset in your research or training pipeline, please reference:

@misc{crowdmind_new_conversation,
  author = {CrowdMind},
  title = {New Conversation: Vision Instruction Dataset},
  year = {2026},
  url = {https://huggingface.co/datasets/CrowdMind/new-conversation}
}

Dataset

Maintained by CrowdMind