---
language:
- en
- it
- es
---

Resource associated to the paper "Extending Large Language Models to Multimodality for non-English Languages".

We provide the following resources:
  - align/: projector alignment dataset of LLaVA translated to Italian and Spanish using madlad
  - test/: contains the MultiInstruct test set formatted in English, as well as the formatted and translated version in Italian and Spanish using MadLad and human written formattings
  - test_conversation/: contains the ImageDialog type 1 subset of the OmniDialog dataset translated to Italian and Spanish using MadLad
  - test_crossmodal-3600/: contains the CrossModal3600 dataset formatted for image captioning and image text selection. It is translated to Italian and Spanish using MadLad and human written formattings
  - train/: contains the MultiInstruct train set formatted in English, as well as the formatted and translated version in Italian and Spanish using MadLad and human written formattings
  - train_conversation/: contains the LLaVA Conv 58k subset of the LLaVA Instruct dataset translated to Italian and Spanish using MadLad

For all these resources, we only provide the text translated using MadLad. You should download the original images following the procedure described by each resource creator.

For more information you can visit the GitHub repository: https://github.com/swapUniba/LVLMs-NonEnglish