VITECH@DravidianLangTech2026: Prompting and LoRA Adaptation for Tamil Abusive Language Detection - A Comparative Study of Open LLMs

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, medium

Summary

This study evaluates four families of open-weight large language models (LLMs)—Gemma, LLaMA, Qwen, and DeepSeek-Distilled—for detecting abusive Tamil text. Researchers assessed models with approximately 2B and 8B parameters using a Tamil dataset from a shared task. The evaluation encompassed two in-context learning settings, zero-shot and few-shot prompting, alongside a parameter-efficient fine-tuning approach utilizing LoRA. Experimental results consistently demonstrated that 8B parameter models surpassed their 2B counterparts, highlighting the advantage of increased model capacity. Among the adaptation techniques, LoRA fine-tuning proved significantly more effective than both zero-shot and few-shot prompting. Google's Gemma-2-9B model, when combined with LoRA fine-tuning, achieved the highest performance across all settings, securing the 12th rank out of 22 submissions with a 0.7959 f1-score.

Key takeaway

For NLP engineers developing abusive language detection systems in low-resource languages like Tamil, you should prioritize LoRA fine-tuning over in-context prompting. Your models will likely benefit from increased capacity; specifically, 8B parameter models consistently outperform 2B versions. Consider Google's Gemma-2-9B with LoRA as a strong baseline, given its superior performance (0.7959 f1-score) in this comparative study. This approach offers a more effective path than traditional prompting for achieving robust detection.

Key insights

LoRA fine-tuning significantly improves open LLM performance for Tamil abusive language detection over prompting.

Principles

• Larger LLMs (8B vs 2B) generally perform better.
• Parameter-efficient fine-tuning (LoRA) excels over in-context learning.
• Model capacity directly impacts performance.

Method

The study evaluated Gemma, LLaMA, Qwen, and DeepSeek-Distilled LLMs (2B/8B) on Tamil abusive text. It compared zero-shot, few-shot prompting, and LoRA fine-tuning on a shared task dataset.

In practice

• Prioritize LoRA fine-tuning for low-resource language tasks.
• Opt for 8B parameter LLMs over 2B for better results.
• Consider Gemma-2-9B with LoRA for Tamil NLP.

Topics

• Tamil NLP
• Abusive Language Detection
• LoRA Fine-tuning
• Large Language Models
• Gemma-2-9B
• Parameter-Efficient Fine-tuning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• System Report for CCL25-Eval Task 5: New Dataset and LoRA-Fine-Tuned Qwen2.5 — Domain-specific datasets and LoRA fine-tuning significantly enhance LLM performance in specialized tasks like classical poetry analysis.
• LoRA: Understanding Low-Rank Adaptation — LoRA approximates full weight updates using low-rank matrices for efficient, latency-free fine-tuning.
• Can AI Save India’s Dying Languages? | Building a Tulu LLM Without Training Data Ft. Prathamesh — Structured prompting and custom romanization enable LLMs to generate low-resource languages without extensive training data.
• How Small Can You Go? LoRA Fine-Tuning 270M-8B Models for Merchant Information Extraction in Financial Transactions — Smaller LoRA-fine-tuned models can achieve near 8B performance for structured extraction, significantly reducing deployment costs.
• LoRA & QLoRA Mastery: The Beginner-to-Advanced Guide to Efficient LLM Fine-Tuning — LoRA and QLoRA enable efficient LLM fine-tuning by adapting a small fraction of parameters, drastically reducing memory and compute.
• A Lightweight LLM Framework for Disaster Humanitarian Information Classification — Parameter-efficient fine-tuning significantly outperforms prompting and RAG for disaster tweet classification, enabling efficient, accurate crisis intelligence.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.