eval-results/leaderboard/mt0xl/mt0xl.json

{"config": {"model_dtype": "", "model_name": "mt0xl", "model_sha": ""}, "results": {"Analytical Talent LSS": {"Acc": 0.28}, "Calculus USS": {"Acc": 0.25}, "Chemistry USS": {"Acc": 0.28}, "Discrete Mathematics USS": {"Acc": 0.26}, "Economy USS": {"Acc": 0.31}, "Geography USS": {"Acc": 0.28}, "Geology USS": {"Acc": 0.25}, "Geometry USS": {"Acc": 0.25}, "History USS": {"Acc": 0.27}, "Logic USS": {"Acc": 0.22}, "Mathematical and Logical Intelligence UPS": {"Acc": 0.21}, "Mathematics LPS": {"Acc": 0.27}, "Mathematics LSS": {"Acc": 0.25}, "Mathematics UPS": {"Acc": 0.27}, "Mathematics USS": {"Acc": 0.25}, "Mathematics and Statistics USS": {"Acc": 0.27}, "Natural Sciences LPS": {"Acc": 0.38}, "Natural Sciences LSS": {"Acc": 0.28}, "Natural Sciences UPS": {"Acc": 0.3}, "Persian Literature LPS": {"Acc": 0.33}, "Persian Literature LSS": {"Acc": 0.3}, "Persian Literature UPS": {"Acc": 0.31}, "Persian Literature USS": {"Acc": 0.26}, "Philosophy USS": {"Acc": 0.33}, "Physics USS": {"Acc": 0.27}, "Probability and Statistics USS": {"Acc": 0.23}, "Psychology USS": {"Acc": 0.35}, "Social Studies LPS": {"Acc": 0.38}, "Social Studies LSS": {"Acc": 0.31}, "Social Studies UPS": {"Acc": 0.36}, "Sociology USS": {"Acc": 0.26}, "Speed and Accuracy UPS": {"Acc": 0.28}, "Theology LPS": {"Acc": 0.51}, "Theology LSS": {"Acc": 0.35}, "Theology UPS": {"Acc": 0.35}, "Theology USS": {"Acc": 0.3}, "Verbal and Linguistic Intelligence UPS": {"Acc": 0.31}, "\u200cBiology USS": {"Acc": 0.26}, "Avg on all tasks": {"Acc": 0.29}, "Avg on all questions": {"Acc": 0.29}}}