Anjir Ahmed Chowdhury
Research Assistant & PhD Candidate @ University of Houston
Large Language Models (LLMs)
đź“° Highlited News
📌Received Bangladesh–Sweden Trust Fund (BSTF) Travel Grant
🧾Passed PhD RCE (Candidacy Exam)
🔬 Ongoing Research
Synthetic Data & Prompt Engineering for LLM Alignment
A. A. Chowdhury, S. Zawad, and F. Yan
Summary : Developed prompt-optimized synthetic data pipelines to improve LLM alignment through instruction design, synthetic dataset creation, and multi-stage fine-tuning. Closely collaborating with Dr. Syed Zawad (IBM Research) as the lead student, contributing extensively across development, experimentation, and evaluation.
Contributions: (i) Designed persona-based prompts to generate synthetic data for reasoning, math, coding, and instruction tasks. (ii) Implemented multi-stage training with iterative SFT updates. (iii) Fine-tuned LLaMA-3.1 8B and Qwen2.5 7B using multi-GPU distributed training. (iv) Used Ollama and vLLM for efficient inference and evaluated models with LM Eval Harness.
PEML: Parameter-efficient Multi-Task Learning with Optimized Continuous Prompts
A. A. Chowdhury, S. Zawad, X. Ma, X. Dong, and F. Yan
Summary: Proposed a unified parameter-efficient framework for multi-task learning, integrating prompt optimization with low-rank adaptation to improve performance across NLP benchmarks. Working closely with Dr. Syed Zawad (IBM Research) and receiving support from Dr. Xiaolong Ma (Argonne National Laboratory). Serving as the lead student overseeing model design, training workflows, and experimental evaluation.
Contributions: (i) Built PEML by integrating PrefixNAS for prompt search with LoRA adaptation. (ii) Evaluated on GLUE, SuperGLUE, and MMLU, achieving up to 6.67% average gains over SOTA. (iii) Created an automated NAS module to optimize prompts and reduce manual tuning. (iv) Achieved efficient resource use with reduced VRAM overhead.
PRENAS: A Provident and Resource Efficient System for Neural Architecture Search
X. Dong, X. Ma, A. A. Chowdhury, S. Zawad, R. Kettimuthu, and F. Yan
Summary: Designed and implemented a resource-efficient NAS system that dynamically allocates computational resources based on architecture scalability and performance prediction to accelerate neural architecture search for LLM fine-tuning and CNN architectures
Contributions: (i) Built the NAS framework with RAY and Optuna for scalable, distributed LLM fine-tuning search. (ii) Integrated distributed training with Accelerate and multi-GPU profiling to measure architecture scalability. (iii) Developed PEFT-NAS-Bench, an open-source benchmark for prefix-tuning NAS, including hyperparameter and scalability profiling.