Enhanced Visual Instruction Tuning with Synthesized Image-Dialogue Data
This repository offers a collection of AI-generated datasets specifically tailored for visual instruction tuning.
conda create -n stablellava python=3.10 -y
conda activate stablellava
pip install --upgrade pip
pip install -e .
-
Download mmbench dev/test set, then put it under
./playground/data/eval/mmbench
-
Set model_path and model_base in
scripts/v1_5/eval/mmbench.sh
For model_path, you can download from google drive
For model_base, we adopt vicuna-13b-v1.5, and you can download from huggingface
Test with single-gpu
CUDA_VISIBLE_DEVICES=0 bash scripts/v1.5/eval/mmbench.sh
-
Submit the results under
./playground/data/eval/mmbench/answers_upload/
to MMBench