* Equally contributing first authors

• Mar-25-24: PALO training and evaluation codes, and pretrained checkpoints are released. 🔥🔥
• Mar-03-24: PALO multi-lingual evaluation dataset is released. Check it out at MBZUAI/multilingual-llava-bench-in-the-wild. 🔥🔥
• Feb-27-24: PALO multi-lingual training dataset is released. Check it out at MBZUAI/palo_multilingual_dataset. 🔥🔥
• Feb-23-24: PALO paper and online demo are released. Code, pretrained models and training/evaluation scripts are coming soon!

In pursuit of more inclusive Vision-Language Models (VLMs), this study introduces a Large Multilingual Multimodal Model called PALO. PALO offers visual reasoning capabilities in 10 major languages, including English, Chinese, Hindi, Spanish, French, Arabic, Bengali, Russian, Urdu, and Japanese, that span a total of ~5B people (65% of the world population).

1. We develop Palo: the first multilingual Large Multimodal Model (LMM), capable of generating responses in 10 languages.
2. We created an extensive multilingual instruction-tuning dataset (~2.1M instructions) by translating LLaVA-Instruct-150K.
3. We train models across three distinct scales i.e., 1.7B, 7B, and 13B parameters to demonstrate the scalability of our training pipeline. The models demonstrate good performance on low-resource languages, e.g., Hindi, Arabic, Bengali, and Urdu, without compromising its high-performance on high-resource languages e.g., English, Chinese, French, and Spanish.

We develop a diverse instruction set (~2.1M instructions) comprising conversations from ten languages. Specifically, 665K instructions from LLaVA-Instruct-665K are used for English, and approximately 150K conversations from LLaVA-Instruct-150K are translated to Chinese, French, Spanish, Russian, Japanese, Arabic, Hindi, Bengali and Urdu using our proposed semi-automated translation pipeline.

📥 Download the Training Dataset: Access our multi-lingual dataset on Hugging Face: MBZUAI/palo_multilingual_dataset.

We also develop a multi-lingual evaluation set to conduct a comprehensive evaluation across various languages. This set is constructed by translating the LLaVA-Bench into all target languages using GPT-4-Turbo, with particular attention to preserving linguistic authenticity and mitigating common issues of automated translations through careful human correction.

📥 Download the Evaluation Dataset: Access our multi-lingual evaluation dataset on Hugging Face: MBZUAI/MBZUAI/multilingual-llava-bench-in-the-wild.

We recommend setting up a conda environment for the project:

conda create --name=palo python=3.10
conda activate palo

git clone https://github.com/mbzuai-oryx/PALO
cd PALO

pip install -r requirements.txt
pip instal flash-attn==2.3.2

export PYTHONPATH="./:$PYTHONPATH"

Please follow the instructions below to run the PALO demo on your local GPU machine.

1. Launch a controller

python palo/serve/controller.py --host 0.0.0.0 --port 10000

2. Launch a gradio web server.

python palo/serve/gradio_web_server.py --controller http:https://localhost:10000 --model-list-mode reload

3. Launch a model worker

python palo/serve/model_worker.py --host 0.0.0.0 --controller http:https://localhost:10000 --port 40000 --worker http:https://localhost:40000 --model-path MBZUAI/PALO-13B

You can launch as many workers as you want, and compare between different model checkpoints in the same Gradio interface. Please keep the --controller the same, and modify the --port and --worker to a different port number for each worker.

1. Prepare data

Please download the annotations from MBZUAI/palo_multilingual_dataset and all images following the below links.

• COCO: train2017
• GQA: images
• OCR-VQA: download script,
• TextVQA: train_val_images
• VisualGenome: part1, part2

After downloading all of them, organize the data as follows in ./playground/data,

data
    ├── coco
    │   └── train2017
    ├── gqa
    │   └── images
    ├── ocr_vqa
    │   └── images
    ├── textvqa
    │   └── train_images
    └── vg
        ├── VG_100K
        └── VG_100K_2
    ├── palo_multilingual_dataset
        ├── palo_multilingual_dataset.json

Please note that all images should be in the .jpg format.

2. Download Pretrained Projection Weights

3. Run Training

# For MobilePALO-1.7B
bash scripts/train/finetune_palo.sh "mtgv/MobileLLaMA-1.4B-Chat" "data/palo_multilingual_dataset/palo_multilingual_dataset.json" <path to palo_1.7B_stage1_mm_projector.bin> "ldpnet" "results/PALO-1.7B" "2" "2e-5"

# For PALO-7B
bash scripts/train/finetune_lora_palo.sh "lmsys/vicuna-7b-v1.5" "data/palo_multilingual_dataset/palo_multilingual_dataset.json" <path to llava-v1.5-mlp2x-336px-pretrain-vicuna-7b-v1.5.bin> "mlp2x_gelu" "results/PALO-7B" "3" "2e-4"

# For PALO-13B
bash scripts/train/finetune_lora_palo.sh "lmsys/vicuna-13b-v1.5" "data/palo_multilingual_dataset/palo_multilingual_dataset.json" <path to llava-v1.5-mlp2x-336px-pretrain-vicuna-13b-v1.5.bin> "mlp2x_gelu" "results/PALO-13B" "3" "2e-4"

Please download PALO multi-lingual evaluation data from MBZUAI/MBZUAI/multilingual-llava-bench-in-the-wild and arrange it as follows,

data
    ├── multilingual-llava-bench-in-the-wild 
        ├── arabic
            ├── question.jsonl
            ├── answers.jsonl
            ├── context.jsonl
        ├── bengali
            ├── question.jsonl
            ├── answers.jsonl
            ├── context.jsonl
        ...
        ...
        ...

Use the following scripts to perform evaluation,

bash scripts/eval/eval_all_languages.sh <path to the trained model> <Output file name> <OpenAI API Key>

  @article{PALO2024,
  title={Palo: A Large Multilingual Multimodal Language Model},
  author={Maaz, Muhammad and Rasheed, Hanoona and Shaker, Abdelrahman and Khan, Salman and Cholakal, Hisham and Anwer, Rao M. and Baldwin, Tim and Felsberg, Michael and Khan, Fahad S.},
  journal={arXiv 2402.14818},
  year={2024},
  url={https://arxiv.org/abs/2402.14818}
}