eval-results/leaderboard/haiku/haiku.json

{"config": {"model_dtype": "", "model_name": "haiku", "model_sha": ""}, "results": {"Analytical Talent LSS": {"Acc": 0.39}, "Calculus USS": {"Acc": 0.29}, "Chemistry USS": {"Acc": 0.3}, "Discrete Mathematics USS": {"Acc": 0.27}, "Economy USS": {"Acc": 0.51}, "Geography USS": {"Acc": 0.5}, "Geology USS": {"Acc": 0.52}, "Geometry USS": {"Acc": 0.32}, "History USS": {"Acc": 0.47}, "Logic USS": {"Acc": 0.4}, "Mathematical and Logical Intelligence UPS": {"Acc": 0.28}, "Mathematics LPS": {"Acc": 0.38}, "Mathematics LSS": {"Acc": 0.36}, "Mathematics UPS": {"Acc": 0.35}, "Mathematics USS": {"Acc": 0.29}, "Mathematics and Statistics USS": {"Acc": 0.36}, "Natural Sciences LPS": {"Acc": 0.71}, "Natural Sciences LSS": {"Acc": 0.46}, "Natural Sciences UPS": {"Acc": 0.63}, "Persian Literature LPS": {"Acc": 0.55}, "Persian Literature LSS": {"Acc": 0.46}, "Persian Literature UPS": {"Acc": 0.53}, "Persian Literature USS": {"Acc": 0.32}, "Philosophy USS": {"Acc": 0.51}, "Physics USS": {"Acc": 0.32}, "Probability and Statistics USS": {"Acc": 0.35}, "Psychology USS": {"Acc": 0.5}, "Social Studies LPS": {"Acc": 0.77}, "Social Studies LSS": {"Acc": 0.57}, "Social Studies UPS": {"Acc": 0.7}, "Sociology USS": {"Acc": 0.44}, "Speed and Accuracy UPS": {"Acc": 0.32}, "Theology LPS": {"Acc": 0.71}, "Theology LSS": {"Acc": 0.61}, "Theology UPS": {"Acc": 0.68}, "Theology USS": {"Acc": 0.45}, "Verbal and Linguistic Intelligence UPS": {"Acc": 0.49}, "\u200cBiology USS": {"Acc": 0.27}, "Avg on all tasks": {"Acc": 0.46}, "Avg on all questions": {"Acc": 0.43}}}