---
language:
  - en
tags:
  - vision
  - multimodal
  - visual-language-model
  - image-understanding
  - vlm
  - parquet
pretty_name: Cleaned VLM Training Dataset
size_categories:
  - 10B<n<100B
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
---


![CrowdMind Cleaned VLM Dataset (1)](https://cdn-uploads.huggingface.co/production/uploads/66e629438ce4fc3270d9a910/0fijwu9HWA6adE1Ad55VA.png)

# Cleaned VLM Training Dataset

**A filtered visual-language dataset for training models to understand images and interpret visual information.**

`CrowdMind/cleaned-vlm` is a cleaned version of [`CrowdMind/categories`](https://huggingface.co/datasets/CrowdMind/categories).

The dataset focuses on examples with greater visual complexity. Entries labeled `low` in the `visual_complexity` field have been removed, leaving only `medium` and `high` complexity examples.

The original columns are preserved.

## Dataset Overview

| Property | Description |
|---|---|
| Dataset | CrowdMind/cleaned-vlm |
| Source | CrowdMind/categories |
| Format | Parquet |
| Modality | Vision and text |
| Complexity levels | Medium, High |
| Removed entries | Low visual complexity |
| Splits | Train |
| Storage size | Approximately 22.1 GB |
| Shards | 56 |

![Colorful Sunburst Category Infographic](https://cdn-uploads.huggingface.co/production/uploads/66e629438ce4fc3270d9a910/FeZKQUpEjnSXreYWn2rco.png)

## What Is This Dataset For?

The goal is to provide a more focused collection of visual-language examples for training and experimentation.

Potential applications include:

- **Visual understanding:** Learning to interpret image content.
- **Visual question answering:** Answering questions using visual evidence.
- **Image reasoning:** Understanding relationships between objects and visual elements.
- **Scene interpretation:** Recognizing and describing complex visual situations.
- **Multimodal instruction tuning:** Connecting natural-language instructions with visual inputs.
- **VLM pretraining and fine-tuning:** Improving visual-language capabilities in multimodal models.

## Data Processing

This dataset was derived from `CrowdMind/categories`.

The cleaning process:

1. Load the original dataset.
2. Inspect the `visual_complexity` field.
3. Remove entries labeled `low`.
4. Retain entries labeled `medium` or `high`.
5. Preserve the existing dataset columns.
6. Export the resulting dataset into Parquet shards.

No additional labels or annotations are claimed to have been generated during this filtering process.

## Visual Complexity

The dataset retains two complexity levels:

| Label | Included | Description |
|---|---|---|
| Low | No | Removed during filtering |
| Medium | Yes | Retained from the source dataset |
| High | Yes | Retained from the source dataset |

The filtering is intended to create a more focused training collection containing examples with greater visual complexity.

## Dataset Structure

The repository stores its training data in Parquet format.

```text
cleaned-vlm/
├── README.md
└── data/
    ├── train-00000-of-00056.parquet
    ├── train-00001-of-00056.parquet
    ├── train-00002-of-00056.parquet
    ├── ...
    └── train-00055-of-00056.parquet
```

The dataset preserves the original columns from `CrowdMind/categories`.

## Intended Use

This dataset is intended for research and development involving:

- Vision-language models (VLMs).
- Multimodal foundation models.
- Visual reasoning systems.
- Image understanding and interpretation.
- Multimodal supervised fine-tuning.
- Experiments involving visual complexity and model performance.

Researchers can use this collection to investigate whether training on more visually complex examples improves a model's ability to understand and reason about images.

## Limitations

- The dataset excludes examples labeled with low visual complexity.
- It does not represent the full distribution of the original source dataset.
- Filtering may introduce a bias toward more visually complex examples.
- The dataset's usefulness depends on the quality and diversity of its original annotations.
- Performance improvements from this filtering approach have not been established by the dataset itself.

## Source Dataset

The original dataset:

[CrowdMind/categories](https://huggingface.co/datasets/CrowdMind/categories)

## Maintainer

**CrowdMind**

Hugging Face: https://huggingface.co/CrowdMind