MSc Thesis Proposal " Impact of Loss Functions in Fine-Tuning LLMs for Improving Sentence Embeddings" By: Bishwadeep Sikder

Thursday, August 8, 2024 - 12:00

The School of Computer Science is pleased to present…

Impact of Loss Functions in Fine-Tuning LLMs for Improving Sentence Embeddings

MSc Thesis Proposal by: Bishwadeep Sikder

Date: Thursday, August 8th 2024

Time:  12:00 pm

Location: Odette Building, Room 110

 

Abstract:

In the pursuit of advancing natural language processing, we present a study involving an extensive analysis of BERT-based language models fine-tuned with diverse loss functions to evaluate tasks in Semantic Textual Similarity (STS) and SentEval (an evaluation toolkit for sentence embeddings). The models were optimized using different combinations of custom loss functions like pairwise cosine similarity, CoSENT, In-Batch Negatives, and Angle optimization. The effectiveness of fine-tuning the models with these loss functions was assessed through Spearman’s Rank correlation coefficient calculation on Semantic Textual Similarity (STS) and classification accuracy on SentEval. By systematically evaluating these loss functions, we aim to develop a framework that yields superior sentence embeddings, facilitating more accurate and robust downstream tasks including sentence similarity and classification. Our approach not only seeks to improve the inherent capabilities of BERT models but also provides valuable insights into the synergistic effects of different loss functions where one mitigates the adverse effects of another, along with control over other fine-tuning parameters like pooling strategy, padding strategy, data collation and tokenization. After extensive experimentation, the results obtained demonstrate significant improvements in benchmark performance.

Keywords: Large Language Models, Fine-Tuning LLMs, BERT, Llama, STS, SentEval
 
Thesis Committee:

Internal Reader: Dr. Hossein Fani              

External Reader: Dr. Abdulkadir A. Hussein          

Advisors: Dr. Alioune Ngom, Dr. Jianguo Lu