The repository follows the directory structure below:
Treble-Counterfactual-VLMs/
├── experiments/
├── lavis/
├── llava/
├── POPE/
└── README.md
Before running the project, please download the required dependencies:
- LAVIS: Clone from LAVIS and place it in the
lavis/
directory. - LLaVA: Clone from LLaVA and place it in the
llava/
directory. - POPE: Clone from POPE and place it in the
POPE/
directory. - Download the COCO train2014 and val2014 datasets, and place them in the following directories:
experiments/data/
├── train2014/
└── val2014/
experiments/test/
contains the test files for LLaVA and InstructBlip.experiments/eval/
contains the GPT evaluation code for assessing the MMHal-Bench performance.experiments/Treble_Counterfactual_utils/
contains the utils for implementing treble counterfactual test-time intervention