Dolci-Think-DPO-7B-decontaminated
Decontaminated version of allenai/Dolci-Think-DPO-7B.
Method: 13-gram overlap detection
Original samples: 150,000
Cleaned samples: 149,513
Removed samples: 487 (0.32%)
MMLU, Ifeval, ARC, COPA, LAMBADA, OpenBookQA, Winogrande, BoolQ, HellaSwag, PIQA, Gsm8k, ALERT, GPQA, MATH, MBPP, HumanEval, SimpleQA, CommonsenseQA, DoNotAnswer, AIME24, LiveCodeBench, MATH500