ATTUN

Running the Model

To run the model, navigate to the folder corresponding to the model variant you wish to execute. Then use the following command:

python3 ../run.py \
    --stratKey stratKey \
    --stratQuery stratQuery \
    --stratComp stratComp \
    --lossEvNumber lossEvNumber \
    --multipleClassifier multipleClassifier \
    --advancedEvidenceLoss advancedEvidenceLoss \
    --learningRate lr \
    --nbEpochs nb_epochs \
    --pathTrain path_to_train_file \
    --pathTest path_to_test_file \
    --nbLabelTU nbLabelTU \
    --stratClaimInput stratClaimInput \
    --pathModel path_to_model (optional) \
    --attaex 0  # Set to 1 to use ATTAEX instead of ATTUN

Parameter Descriptions

Parameter	Description
stratKey	Strategy for aggregating evidence keys. Options: "average", "sepToken", "sepTokenAfter"
stratQuery	Strategy for aggregating queries. Options: "averageClaim", "lastToken", "beforeLastToken", "averageEverything"
stratComp	Comparison strategy between claim and evidence. Options: "ratio", "substraction"
lossEvNumber	Float indicating how much to weight the evidence loss component
multipleClassifier	Boolean: whether to use separate classifiers (True or False)
advancedEvidenceLoss	Float for advanced evidence loss
learningRate	Learning rate (float)
nbEpochs	Number of training epochs (int)
pathTrain	Path to the training dataset
pathTest	Path to the testing dataset
nbLabelTU	Number of target labels
stratClaimInput	Strategy for handling claims. Options: "noClassifier", "lastToken", "averageEverything"
pathModel	(Optional) Path to a pre-trained model
attaex	Set to 1 to use ATTAEX; 0 to use ATTUN

• Provide full or relative paths for pathTrain, pathTest, and pathModel.
• The script supports both ATTUN and ATTAEX variants using the --attaex flag.