This document provides instruction on how to train with Q-Instruct dataset on LLaVA-v1.5 (7B/13B), under the proposed two strategies (mix and after), shown as follows.

Step 0: Pre-requisites

Install LLaVA under the main repository, with flash attention for more efficient training.

git clone
cd LLaVA
pip install -e ".[train]"
pip install flash_attn --no-build-isolation
cd ..

After that, you can conduct low-level visual instruction tuning as follows, under either mix or after strategy.

Step 1: Download Training Datasets

Download Q-Instruct

For the Q-Instruct dataset, download them directly via the following script:

cd LLaVA/playground/data
tar -xf q-instruct-images.tar
rm -f q-instruct-images.tar
cd ../../..

Make sure you have the file structures as follows under LLaVA/playground/data.

├── spaq_koniq
├── livefb_liveitw_aigc
├── cleaned_labels.json

Download Public High-level Instruction Tuning Datasets

If you choose the mix strategy, the high-level datasets also need to be downloaded via the following steps:

  1. Download the annotation of the final mixture our instruction tuning data llava_v1_5_mix665k.json:
wget -P LLaVA/playground/data
  1. Download the images from constituting datasets:

After downloading all of them, organize the high-level data as follows in LLaVA/playground/data,

├── coco
│   └── train2017
├── gqa
│   └── images
├── ocr_vqa
│   └── images
├── textvqa
│   └── train_images
└── vg
    ├── VG_100K
    └── VG_100K_2
  1. Merge the Q-Instruct labels with labels from high-level datasets.
jq -s 'add' LLaVA/playground/data/cleaned_labels.json LLaVA/playground/data/llava_v1_5_mix665k.json > LLaVA/playground/data/mix_cleaned_labels.json

Step 2: Start Training

Please make sure you have enough computational resources before training.

Strategy (a): Mix with High-level Datasets

  • (Pre-requisite) Get Pre-trained Multi-modal Projectors from original LLaVA-v1.5
sh scripts/llava_v1.5/
  • 13B (requires 8x A100 80G), 21h
sh scripts/llava_v1.5/
  • 7B (requires 8x A100 40G), 18h
sh scripts/llava_v1.5/

Strategy (b): After High-level Datasets

  • 13B (requires 8x A100 80G), 3.5h
sh scripts/llava_v1.5/
  • 7B (requires 8x A100 40G), 3h
sh scripts/llava_v1.5/