Assignment-3

This repository contains the python scripts and notebooks for Assignment-3 NLP.

Codebase

The assignment is divided into 3 main sections:

• Pretraining the model(bert-base-uncased) on given dataset (wikitext_raw_2_v1)
• Finetuning the pretrained model for specific tasks (Classification and Question & Answering)
• Evaluation of finetuned models using specified metrics (Classification: Accuracy, Precision, Recall, F1 ; Question-Answering: squad_v2, F1, METEOR, BLEU, ROUGE, exact-match)

To find all the respective code files, refer to the below file structure:

Pre-training

• Pre-training Dataset: wikitext_raw_2_v1
• Pre-training Dataset Eval: wikitext_raw_2_v1_test
• Pre-training Code File: pretraining
• Pre-training Perplexities on Test Dataset: pretrained_perplexities

Fine-tuning

• Classification Task Finetuning: finetuning_classification
• Question & Answering Task Finetuning: finetuning_question_answering

Evaluation

• Evaluation Pre-trained model: eval_pretrain
• Evaluation Classification Finetuned model: eval_classification
• Evaluation Question & Answering Finetuned model: eval_question_answering

Miscellaneous

• Calculating Parameters for models: parameter_calculation

Models

Our pre-trained model is pushed on Hugging Face as: Skratch99/bert-pretrained

Our fine-tuned model for classification is pushed as: Nokzendi/bert_sst2_finetuned

Our fine-tuned model for question & answering is pushed as: Skratch99/finetuned-bert-squadv2

Report

The assignment report can be found here.