LLaVA++: Extending Visual Capabilities with LLaMA-3 and Phi-3

Hanoona Rasheed*, Muhammad Maaz*, Salman Khan, and Fahad Khan

* Equal contributions

Mohamed bin Zayed University of AI (MBZUAI)

---

📢 Latest Updates

• Apr-27-24- Google Colab is released to chat with Phi-3-V-3.8B model, check it out at Google Colab 🔥🔥🔥
• Apr-26-24- Phi-3-V and LLaVA-3-V released: Excited to release the new integration of LLaVA with Phi-3 Mini Instruct and LLaMA-3 Instruct models! Hugging Face 🔥🔥🔥

---

💬 Introduction

This repository enhances the capabilities of the LLaVA 1.5 model, incorporating latest LLMs released this weak🔥, Phi-3 Mini Instruct 3.8B, and LLaMA-3 Instruct 8B.

🏆 Results: Phi-3-V and LLaVA-3-V

Comparison on Benchmarks for Instruction-following LMMS & academic-task-oriented datasets:

Model	MMMU	POPE	MME	MMBench-en	MMBench-cn	SEED-all	SEED-img	SEED-vid	LLaVA-Wild	GQA	Science-QA	Average
LLaVA-v1.5-7B	35.4	85.8	1510.7	64.3	58.3	58.6	66.1	37.3	65.4	62.0	66.8	60.0
LLaVA-v1.5-13B	36.4	85.9	1531.3	67.7	63.6	61.6	68.2	42.7	72.5	63.3	71.6	63.3
LLaMA-3-V-8B	37.1	84.2	1441.1	67.0	57.8	62.8	68.6	41.1	66.2	61.9	78.6	62.5
Phi-3-V-3.8B	37.8	85.6	1470.1	68.2	58.5	62.8	67.7	44.5	70.9	61.7	80.7	63.8

• Average computed excluding MME, and second-best are underlined.

🌟 LLaMA-3-V-8B full fine-tuning results - coming soon!

🤖 Model-Zoo

The following table provides an overview of the available models in our zoo. For each model, you can find links to its Hugging Face page.

Model Name	Hugging Face Link	Summary
LLaVA-Phi-3-mini-4k-instruct-pretrain	Hugging Face	Pretrained on LCS-558K.
LLaVA-Phi-3-mini-4k-instruct-lora	Hugging Face	LoRA weights fine-tuned on LLaVA-Instruct-665K.
LLaVA-Phi-3-mini-4k-instruct	Hugging Face	Merged weights in HuggingFace format.

Model Name	Hugging Face Link	Summary
LLaVA-Meta-Llama-3-8B-Instruct-pretrain	Hugging Face	Pretrained on LCS-558K.
LLaVA-Meta-Llama-3-8B-Instruct-lora	Hugging Face	LoRA weights fine-tuned on LLaVA-Instruct-665K.
LLaVA-Meta-Llama-3-8B-Instruct	Hugging Face	Merged weights in HuggingFace format.

Installation

git clone https://github.com/mbzuai-oryx/LLaVA-pp.git
cd LLaVA-pp
git submodule update --init --recursive

Packages you need to update from LLAVA:

pip install git+https://github.com/huggingface/transformers@a98c41798cf6ed99e1ff17e3792d6e06a2ff2ff3

🚀 Phi-3-V

To integrate Phi-3-V with LLaVA, follow these steps to update the codebase:

# Copy necessary files
cp Phi-3-V/train.py LLaVA/llava/train/train.py
cp Phi-3-V/llava_phi3.py LLaVA/llava/model/language_model/llava_phi3.py
cp Phi-3-V/builder.py LLaVA/llava/model/builder.py
cp Phi-3-V/model__init__.py LLaVA/llava/model/__init__.py
cp Phi-3-V/main__init__.py LLaVA/llava/__init__.py
cp Phi-3-V/conversation.py LLaVA/llava/conversation.py

# Training commands
cp scripts/Phi3-V_pretrain.sh LLaVA/Vi-phi3_pretrain.sh
cp scripts/Phi3-V_finetune_lora.sh LLaVA/Vi-phi3_finetune_lora.sh

Train Phi-3-V

1. Pre-train

cd LLaVA
bash Phi3-V_pretrain.sh

2. Finetune

cd LLaVA
bash Phi3-V_finetune_lora.sh

🚀 LLaMA-3-V

To integrate LLaMA-3-V with LLaVA, follow these steps to update the codebase:

# Copy necessary files
cp LLaMA-3-V/train.py LLaVA/llava/train/train.py
cp LLaMA-3-V/conversation.py LLaVA/llava/conversation.py
cp LLaMA-3-V/builder.py LLaVA/llava/model/builder.py
cp LLaMA-3-V/llava_llama.py LLaVA/llava/model/language_model/llava_llama.py

# Training commands
cp scripts/LLaMA3-V_pretrain.sh LLaVA/LLaMA3-V_pretrain.sh
cp scripts/LLaMA3-V_finetune_lora.sh LLaVA/LLaMA3-V_finetune_lora.sh

Train LLaMA-3-V

1. Pre-train

cd LLaVA
bash LLaMA3-V_pretrain.sh

2. Finetune

cd LLaVA
bash LLaMA3-V_finetune_lora.sh

---

🙏 Acknowledgement

We are thankful to LLaVA, and lmms-eval for releasing their models and code as open-source contributions.

In case if you face any issues or have any questions, please feel free to create an issue or reach out at [email protected] & [email protected].

---