Benchmarking Small Language Models and Small Reasoning Language Models on System Log Severity Classification

This paper benchmarks nine small language models (SLMs) and small reasoning language models (SRLMs) on system log severity classification using real-world `journalctl` data from Linux production servers. It evaluates performance under zero-shot, few-shot, and retrieval-augmented generation (RAG) prompting, revealing strong stratification and significant improvements with RAG for some models. The tiny Qwen3-0.6B notably reaches 88.12% accuracy with retrieval, despite weak performance without it.