Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adapting Safe-for-Work Classifier for Malaysian Language Text: Enhancing Alignment in LLM-Ops Framework

Aisyah Razak, Ariff Nazhan, Kamarul Adha, Wan Adzhar Faiq Adzlan, Mas Aisyah Ahmad, Ammar Azman

TL;DR

This work tackles the scarcity of safe-for-work guardrails for Malay-language text in LLM-Ops by creating the first comprehensive Malay NSFW dataset across nine content categories and developing a tailored classifier. It employs a multi-faceted methodology: manual labeling, zero-shot knowledge distillation from LLMs, centroid and sentiment-based filtering, and active learning to efficiently curate high-quality labels. The best-performing model, mesolitica/malaysian-mistral-191M-MLM, achieves strong accuracy and balanced precision/recall, significantly outperforming baselines such as microsoft/debertav3-base. By releasing the dataset and model publicly, the work advances alignment and safety tooling for Malay-language AI systems and lays groundwork for broader, language-specific LLM-Ops guardrails.

Abstract

As large language models (LLMs) become increasingly integrated into operational workflows (LLM-Ops), there is a pressing need for effective guardrails to ensure safe and aligned interactions, including the ability to detect potentially unsafe or inappropriate content across languages. However, existing safe-for-work classifiers are primarily focused on English text. To address this gap for the Malaysian language, we present a novel safe-for-work text classifier tailored specifically for Malaysian language content. By curating and annotating a first-of-its-kind dataset of Malaysian text spanning multiple content categories, we trained a classification model capable of identifying potentially unsafe material using state-of-the-art natural language processing techniques. This work represents an important step in enabling safer interactions and content filtering to mitigate potential risks and ensure responsible deployment of LLMs. To maximize accessibility and promote further research towards enhancing alignment in LLM-Ops for the Malaysian context, the model is publicly released at https://huggingface.co/malaysia-ai/malaysian-sfw-classifier.

Adapting Safe-for-Work Classifier for Malaysian Language Text: Enhancing Alignment in LLM-Ops Framework

TL;DR

This work tackles the scarcity of safe-for-work guardrails for Malay-language text in LLM-Ops by creating the first comprehensive Malay NSFW dataset across nine content categories and developing a tailored classifier. It employs a multi-faceted methodology: manual labeling, zero-shot knowledge distillation from LLMs, centroid and sentiment-based filtering, and active learning to efficiently curate high-quality labels. The best-performing model, mesolitica/malaysian-mistral-191M-MLM, achieves strong accuracy and balanced precision/recall, significantly outperforming baselines such as microsoft/debertav3-base. By releasing the dataset and model publicly, the work advances alignment and safety tooling for Malay-language AI systems and lays groundwork for broader, language-specific LLM-Ops guardrails.

Abstract

As large language models (LLMs) become increasingly integrated into operational workflows (LLM-Ops), there is a pressing need for effective guardrails to ensure safe and aligned interactions, including the ability to detect potentially unsafe or inappropriate content across languages. However, existing safe-for-work classifiers are primarily focused on English text. To address this gap for the Malaysian language, we present a novel safe-for-work text classifier tailored specifically for Malaysian language content. By curating and annotating a first-of-its-kind dataset of Malaysian text spanning multiple content categories, we trained a classification model capable of identifying potentially unsafe material using state-of-the-art natural language processing techniques. This work represents an important step in enabling safer interactions and content filtering to mitigate potential risks and ensure responsible deployment of LLMs. To maximize accessibility and promote further research towards enhancing alignment in LLM-Ops for the Malaysian context, the model is publicly released at https://huggingface.co/malaysia-ai/malaysian-sfw-classifier.
Paper Structure (17 sections, 12 figures, 3 tables, 2 algorithms)

This paper contains 17 sections, 12 figures, 3 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Taxonomy
  3. Data Source
  4. Social Media
  5. Public Articles
  6. Public Datasets
  7. Methodology
  8. Manual Labeling
  9. Knowledge Distillation of Large Language Models
  10. Centroid Based Filtering
  11. Sentiment Polarity Filtering
  12. Active Learning
  13. Result
  14. Model Comparison
  15. Analysis
  16. ...and 2 more sections

Figures (12)

  • Figure 1: Overall Model Training Framework
  • Figure 2: Manual labeling process using Label Studio LabelStudio.
  • Figure 3: 2D visualization using UMAP
  • Figure 4: Harassment Wordcloud
  • Figure 5: Racist Wordcloud
  • ...and 7 more figures