TLDR - Sometimes, LLMs can generate inappropriate, biased, or factually incorrect responses. This might result in a violation of regulations and can lead to legal issues. LLMGuard is a tool which has the potential to address these LLM risks. LLMGuard can monitor user interactions with an LLM application and flags content against specific behaviours or conversation topics.

--> For video tutorials on top LLM papers, check Kalyan KS YouTube channel

--> For top LLM papers of the week, check the newsletter.

Risks with Large Language Models

• Large Language Models (LLMs) have achieved remarkable performance across various natural language processing (NLP) tasks, showing promise across many domains.

• However, LLMs pose many risks including leaking private information, exhibiting bias, raising ethical concerns and generating harmful content (toxicity, violence, etc.)

Requirement of Safety Measures

• We have techniques to improve LLM safety like Reinforcement Learning with Human Feedback (RLHF). However, constant retraining is necessary in such techniques, making them prohibitive in many cases.

• A much better approach is the post-processing of LLM outputs to apply guardrails, ensuring LLMs stay within acceptable parameters.

Introducing LLMGuard

• LLMGuard is a tool designed to safeguard LLM usage by post-processing both user questions and LLM responses.

• Key Features:

  • Library of Detectors: A modular collection of detectors specialized in identifying undesirable content such as:

    • Racial Bias

    • Violence

    • Blacklisted topics (e.g., politics, religion)

    • Personal Identifiable Information (PII)

    • Toxicity

  • Workflow: LLMGuard passes all inputs and outputs through its detectors. If unsafe content is identified, an automated message replaces the LLM's response.

LLMGuard Detectors (Technical Details)

• Racial Bias Detector: LSTM network trained on a Twitter dataset.

• Violence Detector: Simple count-based vectorization with MLP classifier. Uses the Jigsaw Toxicity Dataset.

• Blacklisted Topics Detector: Fine-tuned BERT model on the 20-NewsGroup Dataset. Can blacklist topics in a plug-and-play fashion.

• PII Detector: Regular expression-based for flexibility in identifying various PII types.

• Toxicity Detector: Uses Detoxify, a BERT-based model trained to detect multiple types of toxicity.

Benefits of LLMGuard

• Provides a flexible and modular tool to detect and prevent harmful LLM outputs.

• Easily customizable with new detectors or changes in blacklisted topics.

• Avoids potential issues with constant LLM retraining.

--> For complete details, refer to the LLMGuard paper.