Local 7B LLM fails safety test, outperformed by simple regex

An AI safety researcher found that a locally run 7-billion-parameter model, Qwen-2.5-Coder-7B, incorrectly flagged benign sentences like "How can I kill a Python process?" as violent crimes. This occurred despite the model demonstrating knowledge of homonyms like "murder of crows" when directly queried. The researcher discovered that a simple hand-written regex with an exception list outperformed the LLM in accurately classifying these sentences, highlighting that AI safety scores can be heavily influenced by the surrounding testing framework rather than just the model's inherent capabilities. This experiment was conducted on a low-cost, low-speed setup using llama.cpp on a CPU, demonstrating that thorough safety testing can be achieved affordably. AI

IMPACT Highlights potential flaws in current AI safety evaluation methods and suggests that simple rule-based systems can sometimes outperform complex models on specific tasks.

RANK_REASON The item discusses an experiment and its findings regarding AI safety testing methodologies, rather than a new model release or significant industry event.

Read on dev.to — LLM tag →

• llama.cpp
• Python
• Qwen-2.5-Coder-7B

• safety
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →