MAGER

1. Install

cd MAGER-GitHub
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

2. Download data

Download MAGER-dataset from Baidu Netdisk:

Link: https://pan.baidu.com/s/1gNF5Wl-nrCiDETUtRYKP-A?pwd=8aku
Code: 8aku

After downloading, put the dataset folder under the project root as data:

MAGER-GitHub/
├── data/
│   ├── politifact/
│   ├── gossipcop/
│   └── mcfend/
├── src/
├── requirements.txt
└── README.md

For example, if the downloaded folder is named MAGER-dataset, run:

cd MAGER-GitHub
mv /path/to/MAGER-dataset ./data

3. Configure LLM API

export AGENT_A_BASE_URL=http://127.0.0.1:8000/v1
export AGENT_G_BASE_URL=http://127.0.0.1:8000/v1
export AGENT_D_BASE_URL=http://127.0.0.1:8000/v1
export AGENT_E_BASE_URL=http://127.0.0.1:8000/v1

export AGENT_A_API_KEY=EMPTY
export AGENT_G_API_KEY=EMPTY
export AGENT_D_API_KEY=EMPTY
export AGENT_E_API_KEY=EMPTY

export AGENT_A_MODEL=qwen3-8b
export AGENT_G_MODEL=qwen3-8b
export AGENT_D_MODEL=qwen3-8b
export AGENT_E_MODEL=qwen3-8b

4. Check data

PYTHONPATH=src/mager python src/mager/check_data.py --dataset politifact
PYTHONPATH=src/mager python src/mager/check_data.py --dataset gossipcop
PYTHONPATH=src/mager python src/mager/check_data.py --dataset mcfend

5. Generate enhanced data if needed

If data/<dataset>/enhanced_dataset_<dataset>_qwen3-8b.jsonl is missing, run:

PYTHONPATH=src/mager python src/mager/run.py preprocess --dataset politifact --limit 400

Change politifact to gossipcop or mcfend for other datasets.

6. Evaluate

PYTHONPATH=src/mager python src/mager/run.py evaluate --dataset politifact --limit 50

7. Evolve meta-paths

PYTHONPATH=src/mager python src/mager/run.py evolve \
  --dataset politifact \
  --evolve_limit 100 \
  --test_limit 200 \
  --max_iterations 100 \
  --population_size 20

Common parameters

• --dataset: dataset name. Choose from politifact, gossipcop, mcfend.
• --limit: number of samples used in evaluation, or number of fake/real samples used separately in preprocessing.
• --evolve_limit: number of samples used for meta-path evolution.
• --test_limit: number of samples used for final evaluation after evolution.
• --max_iterations: number of genetic evolution iterations.
• --population_size: number of meta-path candidates kept in the population.
• --icl_pool_size: size of the candidate pool for in-context example retrieval.
• --top_k: number of retrieved few-shot examples for each query.
• --meta_path_indices: manually select initial meta-paths by index, for example --meta_path_indices 0 1 2.
• --meta_path_json: load selected/evolved meta-paths from a JSON file.
• --no_save: run evaluation without saving result files.