Unsafe-Eval Leaderboard

Balanced Accuracy (BA) and Unsafe F1 average the six benchmarks that mix safe and unsafe prompts. SimpleSafetyTests and HarmBench have only unsafe prompts, so they report only Unsafe Recall. — means a value is unavailable or not applicable. We run local models on a single NVIDIA GeForce RTX 5090 (32 GB) with an AMD Ryzen Threadripper 9970X (32 cores) and 256 GB of RAM.