Models That Know How Evaluations Are Designed Score Safer
Paper • 2605.28591 • Published • 10
🕵️🛡️ Evaluation Meta Knowledge
2026 arXiv preprint. Models fine-tuned on documents describing typical evaluation traits show safer behavior by having increased refusal rates and low
-
-
compass-group-tue/sdf_evaluation_traits
Viewer • Updated • 74.8k • 407 • 1 -
compass-group-tue/sdf_evaluation_traits_15M
Viewer • Updated • 10.6k • 56 -
compass-group-tue/nemotron-traits
Text Generation • Updated • 13 -
compass-group-tue/nemotron-type-hints
Text Generation • Updated • 34 -
compass-group-tue/qwen3-32b-traits
Text Generation • Updated • 52 -
compass-group-tue/qwen3-32b-fineweb
Text Generation • Updated • 22 • 1 -
compass-group-tue/glm-4.7-flash-traits
Text Generation • Updated • 13 -
compass-group-tue/glm-4.7-flash-fineweb
Text Generation • Updated • 18 • 1 -
compass-group-tue/nemotron-6-traits-wo-hr
Text Generation • Updated • 18 -
compass-group-tue/nemotron-6-traits-wo-vs
Text Generation • Updated • 11 -
compass-group-tue/nemotron-traits-15M
Text Generation • Updated • 10 -
compass-group-tue/Nemotron-Traits-Seed-43
Text Generation • Updated • 8 -
compass-group-tue/Nemotron-Traits-Seed-44
Text Generation • Updated • 11 • 1 -
compass-group-tue/deepseek-r1-distill-qwen-32b-traits
Text Generation • Updated • 13 -
compass-group-tue/olmo-3.1-instruct-32b-traits
Text Generation • Updated • 10