Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Protect: Towards Robust Guardrailing Stack for Trustworthy Enterprise LLM Systems

Karthik Avinash, Nikhil Pareek, Rishav Hada

TL;DR

Protect addresses the need for robust guardrails in enterprise LLMs by providing a native multi-modal safety stack for text, image, and audio. It uses four category-specific LoRA adapters trained within an Axolotl-based fine-tuning regime, coupled with a teacher-assisted relabeling pipeline that produces reasoning traces for explainability. Empirical results show state-of-the-art performance across toxicity, sexism, data privacy, and prompt-injection tasks, outperforming open-source and rival proprietary baselines and approaching GPT-4.1 in several dimensions. The work emphasizes production-readiness and auditable safety, and it opens-sources text-only components to spur reproducibility and benchmarking.

Abstract

The increasing deployment of Large Language Models (LLMs) across enterprise and mission-critical domains has underscored the urgent need for robust guardrailing systems that ensure safety, reliability, and compliance. Existing solutions often struggle with real-time oversight, multi-modal data handling, and explainability -- limitations that hinder their adoption in regulated environments. Existing guardrails largely operate in isolation, focused on text alone making them inadequate for multi-modal, production-scale environments. We introduce Protect, natively multi-modal guardrailing model designed to operate seamlessly across text, image, and audio inputs, designed for enterprise-grade deployment. Protect integrates fine-tuned, category-specific adapters trained via Low-Rank Adaptation (LoRA) on an extensive, multi-modal dataset covering four safety dimensions: toxicity, sexism, data privacy, and prompt injection. Our teacher-assisted annotation pipeline leverages reasoning and explanation traces to generate high-fidelity, context-aware labels across modalities. Experimental results demonstrate state-of-the-art performance across all safety dimensions, surpassing existing open and proprietary models such as WildGuard, LlamaGuard-4, and GPT-4.1. Protect establishes a strong foundation for trustworthy, auditable, and production-ready safety systems capable of operating across text, image, and audio modalities.

Protect: Towards Robust Guardrailing Stack for Trustworthy Enterprise LLM Systems

TL;DR

Protect addresses the need for robust guardrails in enterprise LLMs by providing a native multi-modal safety stack for text, image, and audio. It uses four category-specific LoRA adapters trained within an Axolotl-based fine-tuning regime, coupled with a teacher-assisted relabeling pipeline that produces reasoning traces for explainability. Empirical results show state-of-the-art performance across toxicity, sexism, data privacy, and prompt-injection tasks, outperforming open-source and rival proprietary baselines and approaching GPT-4.1 in several dimensions. The work emphasizes production-readiness and auditable safety, and it opens-sources text-only components to spur reproducibility and benchmarking.

Abstract

The increasing deployment of Large Language Models (LLMs) across enterprise and mission-critical domains has underscored the urgent need for robust guardrailing systems that ensure safety, reliability, and compliance. Existing solutions often struggle with real-time oversight, multi-modal data handling, and explainability -- limitations that hinder their adoption in regulated environments. Existing guardrails largely operate in isolation, focused on text alone making them inadequate for multi-modal, production-scale environments. We introduce Protect, natively multi-modal guardrailing model designed to operate seamlessly across text, image, and audio inputs, designed for enterprise-grade deployment. Protect integrates fine-tuned, category-specific adapters trained via Low-Rank Adaptation (LoRA) on an extensive, multi-modal dataset covering four safety dimensions: toxicity, sexism, data privacy, and prompt injection. Our teacher-assisted annotation pipeline leverages reasoning and explanation traces to generate high-fidelity, context-aware labels across modalities. Experimental results demonstrate state-of-the-art performance across all safety dimensions, surpassing existing open and proprietary models such as WildGuard, LlamaGuard-4, and GPT-4.1. Protect establishes a strong foundation for trustworthy, auditable, and production-ready safety systems capable of operating across text, image, and audio modalities.
Paper Structure (16 sections, 1 figure, 11 tables)

This paper contains 16 sections, 1 figure, 11 tables.

Table of Contents

  1. Introduction
  2. Data Collection and Labeling
  3. Audio Synthesis and Augmentation
  4. Teacher-Assisted Annotation and Relabeling
  5. Training Methodology
  6. Base Model and Fine-Tuning Framework
  7. Training Variants
  8. Variant Performance Analysis
  9. Evaluation Study
  10. Evaluation Setup
  11. Results and Analysis
  12. Inference Performance and Deployment Considerations
  13. Conclusion
  14. Appendix
  15. Representative Qualitative Samples
  16. ...and 1 more sections

Figures (1)

  • Figure 1: Audio synthesis pipeline. We form an instruction set from a full-factorial grid over emotion, speaking rate, accent, style, plus seeded pseudorandom lower-order combinations. Text samples are paired with optional voice identities and instructions via seeded pseudorandom sampling, synthesized with CosyVoice 2.0, and optionally augmented with ambient noise at SNRs of 5–20 dB.