Dependencies and package installation guide

The code is written using Python 3.13.2 and requires you to download the following datasets to make use of it all.

Glove embeddings Common Crawl (840B tokens, 2.2M vocab, cased, 300d vectors): https://nlp.stanford.edu/projects/glove/

SentEval for eval.py (follow the guidelines in their github): https://github.com/facebookresearch/SentEval

Then to install all dependencies make use of the environment.yml file and the following command conda env create -f environment.yml and conda activate ACTS

Make sure after activating the conda environment to also additionally pip install SentEval into the environment after cloning their github Do this by running the command: pip install . inside of the SentEval root folder

Structure of the code and file explanation

• models.py, this file contains all 4 encoder models and a simple MLP based classifier model additionally it contains a SNLIModel class that is used to train and evaluate
• train.py, uses the SNLIModel class from models.py you can give the file various arguments when running it to alter the training parameters (explained below)
• utils.py, contains helper functions used by models.py most of these functions make use of the glove embeddings. So make sure to have those installed
• eval.py, file to evaluate a model checkpoint using the SentEval dataset
• demo.ipynb a demonstration on how to use the checkpoints to predict a label based on a NLI task

How to train

You can use the train.py file to train 4 different encoders. The train.py file accepts the following arguments

Command-Line Arguments

The following command-line arguments are available for configuring the model training process:

Argument	Type	Default Value	Description
--seed	int	1234	Random seed for reproducibility.
--encoder_model	str	"uni-lstm"	Encoder model type. Options: uni-lstm, bi-lstm, bi-max-lstm, mean.
--epochs	int	20	Number of training epochs. It is recommended to let it run for a while with a higher value depending on learning rate.
--batch_size	int	64	Size of each batch during training.
--dropout_encoder	float	0.0	Dropout rate for the encoder.
--dropout_classifier	float	0.0	Dropout rate for the classifier.
--lstm_dim	int	512	Dimension of the LSTM layer.
--classifier_dim	int	512	Dimension of the classifier layer.
--class_count	int	3	Number of output classes for classification.
--sgd_lr	float	0.1	Learning rate for Stochastic Gradient Descent (SGD).
--sgd_momentum	float	0.9	Momentum value for SGD.
--decay	float	0.99	Decay rate for the learning rate.
Paths
--outputdir	str	'modelsaves/'	Directory to save model checkpoints and outputs.
--checkpoint_path	str	'none'	Path to a saved checkpoint model (if resuming training).
--outputmodelname	str	'model_output'	Name of the output model file.
--glovepath	str	'glove.840B.300d.txt'	Path to the GloVe word embeddings file (use a pre-trained GloVe file).

Example Usage

python train.py --epochs 50 --encoder_model bi-lstm --outputmodelname bi_lstm_model

How to evaluate with SentEval

You can evaluate model checkpoints using the eval.py file It evaluates on the following transfer tasks 'MR', 'CR', 'MPQA', 'SUBJ'

Example Usage

python eval.py --encoder_model mean --checkpoint_path modelsaves/mean_model.pth

You need to make sure that the encoder_model and checkpoint match. If they do not, the file will fail to evaluate.