🦜 Parrot: Multilingual Visual Instruction Tuning

🎉Introduction • 📰What's New • ☄️Install • 🦜Model • 🔥Train • 🌟Datasets • 🎄MMMB
🔑Evaluation • 📍Quick Start • 👨‍🏫Acknowledgement • 🤗Contact

---

🎉 Introduction

Welcome to Parrot [paper], a novel method that utilizes textual guidance to drive visual token alignment at the language level. Parrot makes the visual tokens condition on diverse language inputs and uses Mixture-of-Experts (MoE) to promote the alignment of multilingual tokens. Moreover, considering the current lack of benchmarks for evaluating multilingual capabilities within the field, we collect and make available a Massive Multilingual Multimodal Benchmark which includes 6 languages, 15 categories, and 12,000 questions, named as MMMB.

If you find Parrot useful for your research and applications, please cite using this BibTeX:

@inproceedings{sun2025parrot,
  title={Parrot: Multilingual Visual Instruction Tuning},
  author={Sun, Hai-Long and Zhou, Da-Wei and Li, Yang and Lu, Shiyin and Yi, Chao and Chen, Qing-Guo and Xu, Zhao and Luo, Weihua and Zhang, Kaifu and Zhan, De-Chuan and others},
  booktitle={ICML},
  year={2025}
}

📰 What's New

• [05/01] 🔥 Parrot is accepted by ICML 2025.
• [08/21] 🔥 We have supported our multilingual MLLM Parrot in VLMEvalKit, now you can evaluate Parrot easily. Welcome to have a try!
• [08/20] 🔥 We have supported MMMB and Multilingual MMBench in VLMEvalKit, now you can use the name MMMB and MTL_MMBench_DEV to obtain the results of 6 langs at the a time. Welcome to have a try!
• [08/02] 🔥 We release the code, inhouse multilingual dataset, benchmark MMMB, and model, welcome to have a try!
• [06/05] 🔥 Parrot is coming! We release the paper!

☄️ Install

Please follow the instructions below to install the required packages.

1. Clone this repository and navigate to Parrot folder

git clone https://github.com/AIDC-AI/Parrot.git
cd Parrot

2. Install Package

conda create -n parrot python=3.10 -y
conda activate parrot
pip install --upgrade pip
pip install -e .

Upgrade to latest code base

git pull
pip install -e . --no-deps

🦜 Model

Parrot is a multilingual multimodal large language model. We provide our fully finetuned models below:

🔥 Train

Parrot is trained in two stages: modality alignment and instruction tuning for multilingual alignment. Each stage's training script is provided in the scripts folder. Before starting the training, ensure you properly set the ROOT variable in the training script. Below are the commands to train Parrot for each stage:

bash scripts/train/pretrain.sh
bash scripts/train/finetune.sh

Hyperparameters

We use a similar set of hyperparameters as Vicuna in finetuning. Both hyperparameters used in pretraining and finetuning are provided below.

1. Pretraining

2. Finetuning

Download Qwen1.5-7B-Chat checkpoints

Our base model Qwen1.5-7B-Chat, which is an instruction-tuned chatbot, can be downloaded from here.

🔎 Datasets

All training datasets are summarized in the Python file located at parrot/train/utils/utils.py. Each dataset contains a collection of samples where each sample consists of text and (optionally) image. The text data is embedded directly within the JSON file, while the image is represented by its filename. This filename refers to the image file located in the image_dir.

We provide the JSON file for each training dataset at Huggingface. The images can be downloaded from their respective sources listed below.

Below is an example of the folder structure. You can alter the folder structure as needed and modify the function name2data in parrot/train/utils/utils.py accordingly.

|-- mllm_datasets
    |-- meta_files
        |-- llava-pretrain-558k.json
        |-- laion-12k.json
        |-- llava-finetune-665k.json
        ...
    |-- images
        |-- llava_pretrain
        |-- sharegpt4v
        |-- laion
        ...

🎄 MMMB

We provide the MMMB benchmark at Huggingface. It contains 6 languages, 15 categories, and 12,000 questions (Following the company's data review, it was identified that some of the data might contain non-compliant information, which could result in the total number of entries in the dataset being slightly fewer than 2,000.) You can download the dataset and use it for your own experiments. We utilize the tsv file to store the dataset, and it is easy to evaluate using the VLMEvalKit.

🔑 Evaluation

We use the VLMEvalKit to evaluate MLLMs.

To evaluate the multilingual capabilities of Parrot, we conduct a comprehensive comparison of it with the state-of-the-art approaches using multilingual benchmarks. Additionally, we compare Parrot with leading models across a range of multimodal tasks. To ensure the reproducibility, we evaluate the models using VLMEvalKit. You can find the evaluation script in VLMEvalKit/run.sh. Before running the script, please replace the paths related to the model and the dataset in the script.

📍 Quick Start

We provide a quick start demo in parrot/deploy/runner.py, which can be used as a template to run Parrot for inference.

1. Before running the demo, please make sure you download the Parrot checkpoint and the Clip checkpoint.
2. Second, you should replace the paths in the runner.py.
3. Finally, run the python file in your system.

👀 Team

This work is a collaborative effort by the MarcoVL team. We would also like to provide links to the following MLLM papers from our team:

• Ovis: Structural Embedding Alignment for Multimodal Large Language Model
• Wings: Learning Multimodal LLMs without Text-only Forgetting

👨‍🏫 Acknowledgement

• LLaVA: the codebase we built upon.
• Qwen1.5-Chat: the LLM backbone we used.
• VLMEvalKit: the evaluation toolkit we used.

🤗 Contact

If there are any questions, please feel free to propose new features by opening an issue or contacting the author: Hai-Long Sun(sunhl@lamda.nju.edu.cn). Enjoy the code!