--- license: other pretty_name: VLM Pretraining task_categories: - image-to-text - visual-question-answering tags: - vision - image - multimodal - vlm - pretraining - image-text - parquet - crowdmind - synthetic-data - image-captioning - visual-language-model size_categories: - 1M` | ## Dataset Contents Each example combines visual information with a descriptive text paragraph. The dataset retains the following information: - **Image:** The visual input associated with the example. - **Category:** Category information providing additional context. - **Text:** Cleaned descriptive caption text derived from assistant-generated descriptions. The original `messages` column has been removed. Instead, captions are represented as descriptive paragraphs. ### Image Token Placement An important characteristic of this dataset is the varied positioning of the `<|image|>` modality token within the descriptive text. This provides examples where image references appear in different positions rather than following one rigid text template. This structure can be useful when experimenting with multimodal sequence formatting and image-text integration. ## Intended Uses This dataset is intended for research and experimentation involving: - Vision Language Model pretraining. - Image-text alignment. - Visual representation learning. - Image caption understanding and generation. - Visual grounding. - Multimodal transformer architectures. - Image-conditioned language modeling. - Experiments with multimodal token placement. ### Recommended Applications **1. VLM Pretraining** Use image-text pairs to train a model to associate visual representations with natural-language descriptions. **2. Multimodal Architecture Experiments** Explore different ways of integrating image embeddings and text tokens into a shared transformer sequence. **3. Visual Grounding** Investigate how models learn relationships between visual content and descriptive language. **4. Training Pipeline Development** Use the dataset to test image loading, multimodal batching, sequence packing, tokenization, and preprocessing pipelines. ## Dataset Structure The repository contains 56 Parquet training shards. ```text vlm-pretraining/ ├── README.md └── data/ ├── train-00000-of-00056.parquet ├── train-00001-of-00056.parquet ├── train-00002-of-00056.parquet ├── ... └── train-00055-of-00056.parquet ``` All shards belong to the training collection. ## Data Processing The dataset is derived from `CrowdMind/new-low`. The preparation process includes: 1. Starting with the source image-text collection. 2. Using descriptive assistant captions. 3. Cleaning captions into readable descriptive paragraphs. 4. Adding category context to the descriptions. 5. Retaining image and category information. 6. Removing the original messages column. 7. Preserving varied placement of the image modality token. 8. Exporting the resulting collection into Parquet shards. ## Limitations Users should consider the following when working with this dataset: - The collection focuses on relatively low visual complexity and should not be treated as a complete representation of real-world visual diversity. - Descriptive captions may contain inaccuracies or omit details present in the original images. - Category context may influence the descriptions and introduce category-related biases. - The dataset is intended for pretraining experiments, not as a standalone benchmark for evaluating general visual intelligence. - Additional filtering, deduplication, and quality assessment may be appropriate for particular training objectives. ## Licensing and Attribution This dataset is derived from: [CrowdMind/new-low](https://huggingface.co/datasets/CrowdMind/new-low) Users should review the source dataset's licensing terms and the licenses applicable to the underlying images before using this collection. The availability of this dataset does not automatically grant additional rights to the underlying images. ## Citation If you use this dataset in your research or training experiments, please cite: ```bibtex @misc{crowdmind_vlm_pretraining, author = {CrowdMind}, title = {VLM Pretraining}, year = {2026}, howpublished = {\url{https://huggingface.co/datasets/CrowdMind/vlm-pretraining}} } ``` ## Related Datasets - [CrowdMind/new-low](https://huggingface.co/datasets/CrowdMind/new-low) - [CrowdMind/cleaned-vlm](https://huggingface.co/datasets/CrowdMind/cleaned-vlm) ## Acknowledgments Created and maintained by **CrowdMind**. The goal of this collection is to make useful multimodal training resources more accessible to researchers and developers experimenting with open Vision Language Models.