GNNVerifier

Graph-based verifier to validate and correct issues in LLM-generated task planning.

Get Started

First, install all the required packages:

conda create -n gnnverifier python=3.8
conda activate gnnverifier
pip install -r requirements.txt

Before running, set your API information in main.sh:

export OPENAI_API_KEY="your_openai_api_key_here"
export OPENAI_API_BASE="your_openai_api_base_here"

Then run the full pipeline (direct prediction, validation, and GNN verification with LLM refinement):

bash main.sh

The pipeline runs in order: direct.py → direct_val.py → main.py.

Configurable Parameters

Edit the variables at the top of main.sh to customize runs:

Variable	Description	Example
LLM_NAME	LLM for prediction & correction	gpt-4o
DATASET	Dataset name	huggingface
LAMBDA_GRAPH	Graph loss weight	2.0
LR	Learning rate	2e-4
LAMBDA_GAP	Gap loss weight	1.5
COST_TAU	Soft target temperature	0.6

Supported datasets: huggingface, multimedia, dailylife, ultratool.