A repository for Neuro-Symbolic Prompting experiments on visual reasoning tasks such as CLEVR and HWF.
Start by creating a Python virtual environment and installing dependencies:
make create_environment
source .venv/bin/activate
make requirements
Ensure your python
command points to a Python 3.10+ interpreter.
First create the data/
directory.
Download from
here
and place under data/HWF
.
Download from here and
place under data/CLEVR_v1.0
.
The main experiments were conducted using the scripts/llm_symbolic_eval.py
script. After setting up the environment as described above, you can run the experiments locally. For example, to run the CLEVR few-shot experiment with logging:
python scripts/llm_symbolic_eval.py \
--dataset clevr \
--log \
--few_shot \
--model gemini-2.0-flash \
--image_before \
--single_turn
For a full list of options, see:
python scripts/llm_symbolic_eval.py --help