Efficient NLP and LLM Adaptation
Knowledge distillation, QLoRA, quantization, adaptive low-rank updates, and memory-efficient training for language models under practical computational constraints.
Read moreKazi Hafiz Md. Asad
M.S. Researcher in Computer Science and Engineering
North South University, Dhaka, Bangladesh
My research focuses on efficient and trustworthy natural language processing, large language model compression and adaptation, graph and hypergraph representation learning, computational biology, and software security. I develop and evaluate methods through controlled experiments, ablation studies, matched baselines, leakage prevention, statistical analysis, and reproducible computational workflows.
Currently preparing for Fall 2027 Ph.D. applications.
29Mparameters in the LastBERT student model
27controlled experiments in current LLM research
9interaction datasets in graph and hypergraph evaluation
Research overview
Efficient machine learning with rigorous empirical evaluation
I study how model efficiency can be improved without weakening the quality of the supporting evidence. My work examines predictive performance together with model size, memory use, runtime, convergence behavior, stability, and generalization.
My current methods include knowledge distillation, low-rank adaptation, quantization, controlled benchmarking, graph and hypergraph modeling, statistical testing, and reproducible experiment design.
Research areas
Primary areas of research
Trustworthy Evaluation and Reproducibility
Matched baselines, ablation studies, leakage control, significance testing, uncertainty analysis, and reproducible experiment pipelines.
Read moreGraph and Hypergraph Learning
Evaluation of pairwise and higher-order representations for protein-complex recovery and other structured learning problems.
Read moreApplied AI for Biology and Security
Efficient and trustworthy learning methods for computational biology, vulnerable-code detection, digital phenotyping, and sustainable AI.
Read moreSelected software project
Attention Is All You Need: Transformer implementation from first principles
This project reconstructs the original encoder-decoder Transformer in PyTorch. It covers token vocabularies, positional encoding, attention masks, multi-head attention, residual connections, normalization, feed-forward layers, training, and inference.
42 implementation steps
PyTorch framework
Seq2Seq training and inference
Open repository
# selected stages in the implementation
01 build_token_to_id_vocab()
12 build_sinusoidal_positional_encoding()
22 scaled_dot_product_attention()
31 assemble_multi_head_attention_forward()
39 encoder_layer_self_attention_sublayer()
42 stack_encoder_layers()Selected publications
Publications and preprints
01
Larger Models Yield Better Results? Streamlined Severity Classification of ADHD-Related Concerns Using BERT-Based Knowledge Distillation
Introduced LastBERT, a 29M-parameter student model with a 3.8x reduction from BERT-base while retaining competitive task performance.
02
Strengthening False Information Propagation Detection: SVM and Text Vectorization in Comparison to BERT
A resource-aware benchmark of BERT and classical SVM pipelines using TF-IDF, Word2Vec, and bag-of-words representations.
03
When Does Higher-Order Representation Improve Protein-Complex Recovery?
A leakage-controlled, matched-k comparison of graph and penalized hypergraph spectral clustering across nine interaction datasets.
Research collaboration
Areas of interest for research collaboration
I am interested in collaborations involving efficient NLP, LLM compression and adaptation, trustworthy evaluation, AI for software engineering, graph learning, computational biology, and resource-aware machine learning.