Optimizing Language Models through Enhanced Fine-Tuning with Data Augmentation Techniques

Paper: Optimizing Language Models through Enhanced Fine-Tuning with Data Augmentation Techniques.pdf

Abstract

Text classification, one of the core tasks of Natural Language Processing (NLP), encounters challenges when evaluating models in out-of-distribution (OOD) contexts. Addressing these challenges requires the application of specialized techniques to enhance model performance. This paper analyzes the efficacy of a fine-tuned iteration of BERT on a custom OOD dataset, utilizing data augmentation techniques to bolster its performance and showcasing the efficacy of this technique. Through a comparative analysis with DistilBERT and GPT-3.5, the paper demonstrates that comparable results can be achieved with a 40% smaller model, emphasizing the potential for efficiency gains without sacrificing performance.

Introduction

Fine-tuning a model a pre-trained model on a downstream task is a common procedure in the NLP space, as it facilitates achieving higher performance with minimal effort. However, one important aspect to consider is that, in real-world scenarios, test data often deviates from the training data distribution. As a result, ensuring that the model exhibits robust performance on datasets with both similar and divergent distributions is crucial.

In this paper, we go over fine-tuning a BERT model on binary classification tasks, testing its performance on a specifically crafted out-of-distribution dataset and discussing the reasons behind the observed decline in the model's effectiveness under these circumstances. Furthermore, the paper encompasses the application of a data augmentation technique involving expanding the training set with out-of-distribution data, followed by a subsequent round of fine-tuning.

We further extend our investigation by applying the previously outlined procedure to DistilBERT, a model that is 40% smaller, highlighting the trade-off between efficiency and performance. To validate the model accuracy, we use GPT-3.5 as a baseline in a zero-shot setting on a small subset of the dataset to verify the model's performance.

The results showcase an enhancement in performance on the out-of-distribution (OOD) dataset after the integration of data augmentation. However, this improvement is accompanied by a comparatively modest decrease in performance on the original dataset. Moreover, the study emphasizes that employing DestilBERT, a smaller model that can be trained 50% faster, enables the preservation of the model's performance in a similar setting.

Getting Started

Installation

Python 3.11.5:

• torch==2.1.0
• datasets==2.14.6
• tqdm==4.66.1
• transformers==4.35.0
• evaluate==0.4.1
• gensim==4.3.2
• nltk==3.8.1

Or you can install them by running:

pip install -r requirements.txt

Files

• main.py: script for fine-tuning and evaluation BERT on the original or transformed dataset.
• main_distilBERT.py: script for fine-tuning and evaluation DistilBERT on the original or transformed dataset.
• utils.py: support script that has all of the transformations to created the out-of-distributions dataset
• word2vec_model.bin: word2vec embeddings used for synonym replacement
• main_GPT.ipynb: Jupyter Notebook for running GPT-3.5 evalaluations of Original (Sample) and Transformed (Sample) datasets as well as BERT and DistilBERt

Predicton files Files within ./pred folder

BERT:

• out_original.txt: Fine-tuned BERT on original dataset
• out_original_transformed.txt: Fine-tuned BERT on transformed dataset
• out_augmented_original.txt: Fine-tuned augmented BERT on original dataset
• out_augmented_transformed: Fine-tuned augmented BERT on transformed dataset
• out_100_original.txt: Fine-tuned BERT predictions on the first 100 rows of the original dataset
• out_augmented_100_transformed.txt: Fine-tuned augmented BERT predictions on the first 100 rows of the transformed dataset

DistilBERT:

• out_distilbert_original.txt: Fine-tuned DistilBERT on original dataset
• out_distilbert_original_transformed.txt: Fine-tuned DistilBERT on transformed dataset
• out_distilbert_augmented_original.txt: Fine-tuned augmented DistilBERT on original dataset
• out_distilbert_augmented_transformed.txt: Fine-tuned augmented DistilBERT on transformed dataset
• out_distilbert_100_original.txt: Fine-tuned DistilBERT predictions on the first 100 rows of the original dataset
• out_distilbert_augmented_100_transformed.txt: Fine-tuned augmented DistilBERT predictions on the first 100 rows of the transformed dataset

GPT3.5 (zero-shot):

• gpt_out_original.txt: prediction on the first 100 rows of the original dataset
• gpt_out_transformed.txt: prediction on the first 100 rows of the transformed dataset

CARC Output Files ./CARC_output/: contain all of CARC outputs for each training and evaluation that were executed

Usage

Fine-Tuning and Evaluating on Original Dataset

python3 main.py --train --eval

Outputs:

• out/: model tensors
• out_original.txt: predictions

python3 main_distilBERT.py --train --eval

Outputs:

• out_distilbert/: model tensors
• out_distilbert_original.txt: predictions

Fine-Tuning and Evaluating on Transformed Dataset

python3 main.py --train_augmented --eval_augmented

Outputs:

• out_augmented/: model tensors
• out_augmented_original.txt: predictions

python3 main_distilBERT.py --train_augmented --eval_augmented

Outputs:

• out_distilbert_augmented/: model tensors
• out_distilbert_augmented_original.txt: predictions

Evaluations

# Evaluation original BERT model on transformed data
python3 main.py --eval_augmented --model_dir ./out

# Evaluation augmented BERT model on original data
python3 main.py --eval_augmented --model_dir ./out_augmented

# Evaluation original DistilBERT model on transformed data
python3 main_distilBERT.py --eval_augmented --model_dir ./out_distilbert

# Evaluation augmented DistilBERT model on original data
python3 main_distilBERT.py --eval_augmented --model_dir ./out_distilbert_augmented