Leaderboards¶
One leaderboard per trait, models ordered by value, with each 95% interval and the rank range the paired comparisons allow.
- General guard · setting-invariant classifier
- World guard · measured on 1 setting
- Prefill, 1K prompt · measured on one GPU
- Prefill, 4K prompt · measured on one GPU
- Generation, 1K prompt · measured on one GPU
- Generation, 4K prompt · measured on one GPU
- VRAM at 8K context · one agent, 8K context, q8_0 KV cache