- dataset: id: FutureMa/EvasionBench task_id: evasion_bench value: 77.83 date: '2026-10-05' source: url: https://huggingface.co/datasets/FutureMa/EvasionBench name: EvasionBench user: SeaWolf-AI notes: 'Own evaluation with inspect-ai, task as defined in the dataset eval.yaml (split train, 16,726 samples, shuffle_choices, multiple_choice template, choice scorer, accuracy). Served checkpoint: FINAL-Bench/Darwin-180B-RSI-R3 (BF16) on vLLM; temperature 1.0, top_p 0.95, max_tokens 8192, default chat template; 13,018 of 16,726 correct, single run.'