ViTucano-2b8-v1

Model Summary

ViTucano is 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. ViTucano integrates visual understanding with linguistic capabilities, creating a tool for multimodal tasks (e.g., image captioning, visual question answering, etc.).

Details

Architecture: TinyLlavaForConditionalGeneration
Vision Tower: google/siglip-so400m-patch14-384
Size: 2,882,351,680 parameters
Context length: 4096 tokens
Dataset:
Language: Portuguese
GPU: 8 NVIDIA A40
Training time: ~ 27 hours
Emissions: 25.4 KgCO2 (Germany)
Total energy consumption: 66 kWh

This repository has the source code used to train this model.

Intended Uses

The primary intended use of the ViTucano models is to serve as foundations for research and development involving native Portuguese foundation models. You may also fine-tune and adapt ViTucano models for deployment if your use follows the Apache 2.0 license. If you decide to use the ViTucano models as a basis for your fine-tuned model, please conduct your own risk and bias assessment.

Out-of-scope Use

ViTucano models are not intended for deployment. They are not an out-of-the-box product and should not be used for human-facing interactions.
ViTucano models are for the Portuguese language only and are unsuitable for image-to-text generation tasks in other languages.
ViTucano models have not been fine-tuned for any specific downstream task.

Basic usage

⚠️Using ViTucano models through the transformers library requires executing remote code (trust_remote_code=True). The executed files are configuration.py and modeling_tinyllava_tucano.py, both available in this repository.⚠️

Run inference using tinyllava

from tinyllava.eval.run_tiny_llava import eval_model

model_path = "TucanoBR/ViTucano-2b8-v1"
prompt = "Quais os principais elementos dessa imagem?"
image_file = "https://raw.githubusercontent.com/Nkluge-correa/TinyLLaVA_Factory/refs/heads/main/assets/sample.jpg"
conv_mode = "llama"

args = type('Args', (), {
    "model_path": model_path,
    "model": None,
    "query": prompt,
    "conv_mode": conv_mode,
    "image_file": image_file,
    "sep": ",",
    "temperature": 0,
    "top_p": None,
    "num_beams": 1,
    "max_new_tokens": 512
})()

eval_model(args)

Run inference using transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_path = "TucanoBR/ViTucano-2b8-v1"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = AutoModelForCausalLM.from_pretrained(
  model_path, 
  #torch_dtype=torch.bfloat16, # for optimized inference  🚀
  #attn_implementation="flash_attention_2" # for optimized inference  🚀
  trust_remote_code=True)
model.to(device)

tokenizer = AutoTokenizer.from_pretrained(model_path)
prompt = "Quais os principais elementos dessa imagem?"
image_file="https://raw.githubusercontent.com/Nkluge-correa/TinyLLaVA_Factory/refs/heads/main/assets/sample.jpg"
output_text, _ = model.chat(prompt=prompt, image=image_file, tokenizer=tokenizer)

print(output_text)

Limitations

Like almost all other multimodal language models trained on large datasets scraped from the web, the ViTucano models show behavior that does not make them an out-of-the-box solution to many real-world applications, especially those requiring factual, reliable, and nontoxic text generation. ViTucano models are all subject to the following:

Hallucinations: ViTucano models may generate misleading or entirely false information when interpreting or describing visual inputs, leading to hallucinations that could be mistaken as accurate observations or factual statements.
Biases and Toxicity: ViTucano models inherit social and historical stereotypes in the training data. These biases can manifest in harmful, offensive, or misleading descriptions or analyses of visual or textual content.
Unreliable Visual Interpretations: ViTucano models may produce inaccurate interpretations of visual elements, including objects, scenes, or text within images. Such outputs should not be considered reliable without human verification.
Multimodal Language Limitations: While ViTucano models are optimized for Portuguese, handling multilingual visual and textual contexts may lead to errors, misinterpretations, or inadequate responses, especially with non-Portuguese content.
Repetition and Irrelevant Details: ViTucano models can exhibit repetitive response patterns or generate verbose descriptions unrelated to the given visual or textual input, particularly under specific hyperparameter configurations.

Hence, even though our models are released with a permissive license, we urge users to perform their risk analysis before using them for real-world applications.

Cite as 🤗

ViTucano

@misc{correa2025vitucano,
    author={Corr{\^e}a, Nicholas Kluge and Sen, Aniket and Falk, Sophia and Fatimah, Shiza},
    title={{ViTucano: A Portuguese Vision Assistant}},
    year=2025,
    howpublished={\url{https://huggingface.co/TucanoBR/ViTucano-2b8-v1}},
    doi={10.57967/hf/4530},
    publisher={{Hugging Face}}
}

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}, 
}

@article{correa2025tucanoadvancingneuraltext,
    title={{Tucano: Advancing Neural Text Generation for Portuguese}},
    author={Corr{\^e}a, Nicholas Kluge and Sen, Aniket and Falk, Sophia and Fatimah, Shiza},
    journal={Patterns},
    publisher={Elsevier},
    year={2025},
    doi={10.1016/j.patter.2025.101325},
    url={https://doi.org/10.1016/j.patter.2025.101325},
    issn={2666-3899}
}

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.

License

ViTucano is licensed under the Apache License, Version 2.0. For more details, see the LICENSE file.

TucanoBR/ViTucano-2b8-v1