Statement & publications
Three threads I am pulling on — and the publications that document each.
Thread 01
Probing
What does a frozen transformer already know? Linear probes as cheap diagnostics for Bangla.
Thread 02
Adaptation
When probes fail — when do we need LoRA, and how should the two be combined?
Thread 03
Evaluation
Building Bangla benchmarks where current ones fall short — medical VQA, hate speech.
“How much of a pre-trained model’s competence transfers, where does transfer fail, and what is the cheapest intervention that recovers the gap?”
Evaluating Large Vision Language Models on Bangla Medical Visual Question Answering
Rafid Ahmed · Intesar Tahmid · Mir Sazzat Hossain · Tasnimul Hossain Tomal · Md Fahim · Md Farhad Alam
A rigorous evaluation of large vision-language models on Bangla medical VQA — exposing systematic failures in clinical terminology translation, visual grounding, and medical reasoning for low-resource languages.
Decoupling Representation for Fine-tuning with Linear Probing and LoRA on Robust Downstream Adaptation
Tasnimul Hossain Tomal · Intesar Tahmid · Rafid Ahmed · Mir Sazzat Hossain · Md Fahim · Md Farhad Alam
We propose decoupling representation learning from task adaptation: a linear probe locks general features, LoRA tunes the rest. The combination is robust under distribution shift.
LP-FT-LORA: A Three-Stage PEFT Framework for Efficient Domain Adaptation in Bangla NLP Tasks
Tasnimul Hossain Tomal · Anam Borhan Uddin · Intesar Tahmid · Mir Sazzat Hossain · Md Fahim · Md Farhad Alam
A three-stage PEFT framework — linear probe initializes, full fine-tuning adapts, LoRA efficiently specializes — yielding consistent gains on Bangla NLP tasks with far fewer trainable parameters.
PentaML @BLP Shared Task 1: Linear Probing of Pre-trained Transformer based Models for Bangla Hate Speech Detection
Intesar Tahmid · Rafid Ahmed · Md Mahir Jawad · Anam Borhan Uddin · Md Fahim · Md Farhad Alam Bhuiyan
Submission to BLP-2025 Shared Task 1. Demonstrates that frozen-encoder probes are a strong, sample-efficient baseline for Bangla hate-speech classification.
How Good LLMs Are at Answering Bangla Medical Visual Questions? Dataset and Benchmarking
Rafid Ahmed · Intesar Tahmid · Mir Sazzat Hossain · Tasnimul Hossain Tomal · Md Fahim · Md Farhad Alam
Benchmarks LLM performance on Bangla medical visual questions, contributing a novel dataset. Reveals that frontier models struggle significantly on Bangla clinical terminology and visual grounding.
❦ ❦ ❦
Currently reading
Service
Statement of research available on request.