Reproduce “<think> So let's replace this phrase with insult... </think> Lessons learned from generation of toxic texts with LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sergey Pletenev, Daniil Moskovskiy, Alexander Panchenko
Research from AIRI and Skoltech evaluated the utility of Large Language Models for generating synthetic toxic text, revealing that LLM-generated data consistently lacked lexical diversity, which negatively impacted the performance of downstream text detoxification models.
Key insights
- LLM-generated toxic text exhibits a significant 'lexical diversity gap,' characterized by a small, repetitive vocabulary of insults and profanities.
- This lack of diversity in synthetic training data leads to detoxification models that generalize poorly to the broader, nuanced spectrum of human-generated toxic language.
- Simply increasing the scale of the LLM (e.g., from 8B to 72B parameters) did not inherently resolve the problem of repetitive toxic output or improve lexical diversity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.