This repository contains the source code and results for the experiments presented in The Impact of Parameter Count on Sarcasm Detection Using BERT-Based Models.
Requirements: Python 3.13 (or a compatible Python version)
Create and activate a virtual environment (optional but recommended). Then install the required dependencies:
pip install --upgrade pip
pip install -r requirements.txtTo run an experiment, you can use the following example and adjust parameters as needed. A simple bash script for running experiments can be found under scripts/run-experiment.sh.
python -m src.main \
--seed 1 \
--dataset news_headlines \
--pretrained-model-name prajjwal1/bert-tiny \
--classification-head bert_single_token_attention \
--dropout-p 0.1 \
--train-test-split 0.7 \
--batch-size 32 \
--learning-rate 2e-5 \
--max-grad-norm 1.0 \
--num-train-epochs 5 \
--experiment-name my-experiment \
--save-modelThe main experiments were conducted using the following scripts:
Grid Search (scripts/grid-search.sh): Evaluates a range of learning rates and batch sizes for five BERT-based models on three datasets. Results are saved to results/logs/grid_search_*/.
Multi-Seed Evaluation (scripts/evaluation.sh): Repeats each model–dataset combination ten times (seeds 1–10) using the best batch size and learning rate selected from the grid search to obtain robust evaluation results. Results are saved to results/logs/evaluation_*/.
Interactive terminal app for exploring trained models with real-time predictions and token saliency visualization.
python inference-app.py --path <path-to-model> --device <device>Arguments:
--path: Path to trained model file (required)--device: PyTorch device (default:cpu)
The app loads the model configuration from the experiment logs.
Trained models are available on Zenodo: https://doi.org/10.5281/zenodo.18331394