Benchmarks and write-ups from building pwner.
Two open security LLMs against the general models they were trained from, on one desktop machine. Plus two broken GGUFs and a refusal scorer that miscounts.