ViTucano-Pretrain

Dataset Summary

ViTucano-Pretrain is a translation of the original liuhaotian/LLaVA-Pretrain, obtained via Google's translation API. LLaVA Visual Instruct Pretrain LCS-558K is a subset of the LAION/CC/SBU dataset, filtered with a more balanced concept coverage distribution. This dataset was used to train the ViTucano, our first attempt at creating a vision assistant natively pretrained in Portuguese. ViTucano is built on top of the Tucano series using the TinyLLaVA Factory.

Supported Tasks and Leaderboards

This dataset can be utilized for tasks involving language modeling and visual instruction tunning.

Languages

Portuguese.

Dataset Structure

Data Instances

The dataset consists of the following features:

• id: an identifier (name of the respective file) for that image.
• image: the path to the file in the original folder configuration.
• conversations: a list of dictionaries, where each dictionary represents a message or an entry in a conversation.
• blip_caption: the original BLIP caption.
• url: the url of the corresponding image.

Data Fields

{
    "id": "004539375",
    "image": "train/00453/004539375.jpg",
    "conversations": [
        {
            "from": "human",
            "value": "Renderize um resumo claro e conciso da foto.\n<image>"
        },
        {
            "from": "gpt",
            "value": "Selecione móveis de luxo 3 - colchão de espuma de memória de gel de polegada"
        }
    ],
    "blip_caption": "Selecione móveis de luxo 3 - colchão de espuma de memória de gel de polegada",
    "url": "http://ec1.ostkcdn.com/images/products/8111140/P15459545.jpg"
}

Data Splits

Available splits are train.

To use this dataset, you will need to download both the data-pretraining.json and images.zip files available in this folder:

wget https://huggingface.co/datasets/TucanoBR/ViTucano-Pretrain/resolve/main/data-pretraining.json
wget https://huggingface.co/datasets/TucanoBR/ViTucano-Pretrain/resolve/main/images.zip

You can also do this via the huggingface_hub library:

from huggingface_hub import snapshot_download

snapshot_download(repo_id="ViTucano-Pretrain", repo_type="dataset")

Unzip the images in a way that you get this folder structure (e.g., unzip images.zip -d "path/to/train"):

├── train
    ├── 00000
    ├── 00001
    ├── 00002
    └── etc ...

Done! The data is ready to train your projector.

Dataset Creation

Curation Rationale

This dataset is a translation of the original liuhaotian/LLaVA-Pretrain obtained via Google's translation API.

Source Data

Who are the source language producers?

All text samples translated from English to Portuguese.

Annotations

Annotation process

Read this dataset card for more information.

Who are the annotators?

Read this dataset card for more information.

Considerations for Using the Data

Warning: This dataset may contain NSFW (Not Safe For Work) content, including explicit images and text captions with offensive/sensitive language.

Other Known Limitations

This dataset has has been translated using translation engines, potentially resulting in corrupted samples. While useful for quickly converting text between languages, translation engines often struggle with accurately preserving the syntax, semantics, and context of certain languages.

Additional Information

Dataset Curators

Nicholas Kluge Corrêa.

Licensing Information

Users of this dataset must comply with license of CC-3M and BLIP (if you use their synthetic caption).

Creative Commons Attribution 4.0 International; and it should abide by the policy of OpenAI.

Citation Information

ViTucano

@misc{correa20204vitucano,
    author={Corr{\^e}a, Nicholas Kluge and Sen, Aniket and Falk, Sophia and Fatimah, Shiza},
    title={{ViTucano: A Portuguese Vision Assitant}},
    year=2024,
    howpublished = {\url{https://huggingface.co/TucanoBR}},
}

Tucano

@misc{correa2024tucanoadvancingneuraltext,
      title={{Tucano: Advancing Neural Text Generation for Portuguese}}, 
      author={Corr{\^e}a, Nicholas Kluge and Sen, Aniket and Falk, Sophia and Fatimah, Shiza},
      year={2024},
      eprint={2411.07854},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2411.07854}, 
}

TinyLLaVA Factory

@article{jia2024tinyllava,
  title={TinyLLaVA Factory: A Modularized Codebase for Small-scale Large Multimodal Models},
  author={Jia, Junlong and Hu, Ying and Weng, Xi and Shi, Yiming and Li, Miao and Zhang, Xingjian and Zhou, Baichuan and Liu, Ziyu and Luo, Jie and Huang, Lei and Wu, Ji},
  journal={arXiv preprint arXiv:2405.11788},
  year={2024}
}

LLaVA

@misc{liu2023llava,
      title={Visual Instruction Tuning}, 
      author={Liu, Haotian and Li, Chunyuan and Wu, Qingyang and Lee, Yong Jae},
      publisher={NeurIPS},
      year={2023},
}

Aknowlegments

We gratefully acknowledge the granted access to the Marvin cluster hosted by University of Bonn along with the support provided by its High Performance Computing & Analytics Lab.

Contributions

If you want to contribute, contact me at [email protected]!