GlitchTokenAvoidance

This Approach Presents a first way of blackbox mitigating malicious or unexpected behavior of LLMs that are error-prone to Glitch Tokens

What is the context here?

Glitch Tokens are phenomena that occur in models that use a tokenizer to initiate the embedding process of a model execution. These tokens can potentially harm the system and lead to false or unexpected results. To avoid the accidental usage of these tokens in a model where these glitch tokens are present and not removable, this technique comes to use.

What exactly is done?

Since the deployed model offers no ad-hoc mitigation, a sanitization step is introduced to ensure safe model usage. The sanitization is done by a variation of granularity levels. Which granularity is optimal for the problem is open to discussion and context. This application will help understanding the differences between these different strategies.

How to use

1. Pick a strategy by selecting the token-, word-, sentence- or prompt-based approach.
2. Insert your prompt and receive a model response. If the prompt contains any glitch tokens, the prompt will be automatically adjusted based on the chosen strategy.
3. process the returned string in a desired way.

Example

from GlitchTokenAvoidance import GlitchTokenAvoidance as gta

Question = "Where do I apply for a dissertation?"
Model = "Llama3.2"
Replacement_Model = "mistral"

result = gta.ModelRequest_Sentence_based(Question, Model, Replacement_Model)
print(result)