@inproceedings{tomal-etal-2025-lp,
title = "{LP}-{FT}-{L}o{RA}: A Three-Stage {PEFT} Framework for Efficient Domain Adaptation in {B}angla {NLP} Tasks",
author = "Tomal, Tasnimul Hossain and
Uddin, Anam Borhan and
Tahmid, Intesar and
Hossain, Mir Sazzat and
Fahim, Md and
Bhuiyan, Md Farhad Alam",
editor = "Alam, Firoj and
Kar, Sudipta and
Chowdhury, Shammur Absar and
Hassan, Naeemul and
Prince, Enamul Hoque and
Tasnim, Mohiuddin and
Rony, Md Rashad Al Hasan and
Rahman, Md Tahmid Rahman",
booktitle = "Proceedings of the Second Workshop on Bangla Language Processing (BLP-2025)",
month = dec,
year = "2025",
address = "Mumbai, India",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.banglalp-1.17/",
pages = "212--222",
ISBN = "979-8-89176-314-2",
abstract = "Adapting large pre-trained language models (LLMs) to downstream tasks typically requires fine-tuning, but fully updating all parameters is computationally prohibitive. Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) reduce this cost by updating a small subset of parameters. However, the standard approach of jointly training LoRA adapters and a new classifier head from a cold start can lead to training instability, as the classifier chases shifting feature representations. To address this, we propose LP-FT-LoRA, a novel three-stage training framework that decouples head alignment from representation learning to enhance stability and performance. Our framework first aligns the classifier head with the frozen backbone via linear probing, then trains only the LoRA adapters to learn task-specific features, and finally performs a brief joint refinement of the head and adapters. We conduct extensive experiments on five Bangla NLP benchmarks across four open-weight compact transformer models. The results demonstrate that LP-FT-LoRA consistently outperforms standard LoRA fine-tuning and other baselines, achieving state-of-the-art average performance and showing improved generalization on out-of-distribution datasets."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="tomal-etal-2025-lp">
<titleInfo>
<title>LP-FT-LoRA: A Three-Stage PEFT Framework for Efficient Domain Adaptation in Bangla NLP Tasks</title>
</titleInfo>
<name type="personal">
<namePart type="given">Tasnimul</namePart>
<namePart type="given">Hossain</namePart>
<namePart type="family">Tomal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anam</namePart>
<namePart type="given">Borhan</namePart>
<namePart type="family">Uddin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Intesar</namePart>
<namePart type="family">Tahmid</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mir</namePart>
<namePart type="given">Sazzat</namePart>
<namePart type="family">Hossain</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Md</namePart>
<namePart type="family">Fahim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Md</namePart>
<namePart type="given">Farhad</namePart>
<namePart type="given">Alam</namePart>
<namePart type="family">Bhuiyan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-12</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Second Workshop on Bangla Language Processing (BLP-2025)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Firoj</namePart>
<namePart type="family">Alam</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sudipta</namePart>
<namePart type="family">Kar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shammur</namePart>
<namePart type="given">Absar</namePart>
<namePart type="family">Chowdhury</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Naeemul</namePart>
<namePart type="family">Hassan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Enamul</namePart>
<namePart type="given">Hoque</namePart>
<namePart type="family">Prince</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohiuddin</namePart>
<namePart type="family">Tasnim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Md</namePart>
<namePart type="given">Rashad</namePart>
<namePart type="given">Al</namePart>
<namePart type="given">Hasan</namePart>
<namePart type="family">Rony</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Md</namePart>
<namePart type="given">Tahmid</namePart>
<namePart type="given">Rahman</namePart>
<namePart type="family">Rahman</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Mumbai, India</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-314-2</identifier>
</relatedItem>
<abstract>Adapting large pre-trained language models (LLMs) to downstream tasks typically requires fine-tuning, but fully updating all parameters is computationally prohibitive. Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) reduce this cost by updating a small subset of parameters. However, the standard approach of jointly training LoRA adapters and a new classifier head from a cold start can lead to training instability, as the classifier chases shifting feature representations. To address this, we propose LP-FT-LoRA, a novel three-stage training framework that decouples head alignment from representation learning to enhance stability and performance. Our framework first aligns the classifier head with the frozen backbone via linear probing, then trains only the LoRA adapters to learn task-specific features, and finally performs a brief joint refinement of the head and adapters. We conduct extensive experiments on five Bangla NLP benchmarks across four open-weight compact transformer models. The results demonstrate that LP-FT-LoRA consistently outperforms standard LoRA fine-tuning and other baselines, achieving state-of-the-art average performance and showing improved generalization on out-of-distribution datasets.</abstract>
<identifier type="citekey">tomal-etal-2025-lp</identifier>
<location>
<url>https://aclanthology.org/2025.banglalp-1.17/</url>
</location>
<part>
<date>2025-12</date>
<extent unit="page">
<start>212</start>
<end>222</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T LP-FT-LoRA: A Three-Stage PEFT Framework for Efficient Domain Adaptation in Bangla NLP Tasks
%A Tomal, Tasnimul Hossain
%A Uddin, Anam Borhan
%A Tahmid, Intesar
%A Hossain, Mir Sazzat
%A Fahim, Md
%A Bhuiyan, Md Farhad Alam
%Y Alam, Firoj
%Y Kar, Sudipta
%Y Chowdhury, Shammur Absar
%Y Hassan, Naeemul
%Y Prince, Enamul Hoque
%Y Tasnim, Mohiuddin
%Y Rony, Md Rashad Al Hasan
%Y Rahman, Md Tahmid Rahman
%S Proceedings of the Second Workshop on Bangla Language Processing (BLP-2025)
%D 2025
%8 December
%I Association for Computational Linguistics
%C Mumbai, India
%@ 979-8-89176-314-2
%F tomal-etal-2025-lp
%X Adapting large pre-trained language models (LLMs) to downstream tasks typically requires fine-tuning, but fully updating all parameters is computationally prohibitive. Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) reduce this cost by updating a small subset of parameters. However, the standard approach of jointly training LoRA adapters and a new classifier head from a cold start can lead to training instability, as the classifier chases shifting feature representations. To address this, we propose LP-FT-LoRA, a novel three-stage training framework that decouples head alignment from representation learning to enhance stability and performance. Our framework first aligns the classifier head with the frozen backbone via linear probing, then trains only the LoRA adapters to learn task-specific features, and finally performs a brief joint refinement of the head and adapters. We conduct extensive experiments on five Bangla NLP benchmarks across four open-weight compact transformer models. The results demonstrate that LP-FT-LoRA consistently outperforms standard LoRA fine-tuning and other baselines, achieving state-of-the-art average performance and showing improved generalization on out-of-distribution datasets.
%U https://aclanthology.org/2025.banglalp-1.17/
%P 212-222
Markdown (Informal)
[LP-FT-LoRA: A Three-Stage PEFT Framework for Efficient Domain Adaptation in Bangla NLP Tasks](https://aclanthology.org/2025.banglalp-1.17/) (Tomal et al., BanglaLP 2025)
ACL