nanchennn/Adv_MATH500

--- dataset_info: - config_name: easy features: - name: text dtype: string splits: - name: train num_bytes: 154840 num_examples: 500 download_size: 76937 dataset_size: 154840 - config_name: hard features: - name: text dtype: string splits: - name: train num_bytes: 442269 num_examples: 500 download_size: 197834 dataset_size: 442269 - config_name: medium features: - name: text dtype: string splits: - name: train num_bytes: 195315 num_examples: 500 download_size: 98901 dataset_size: 195315 configs: - config_name: easy data_files: - split: train path: easy/train-* - config_name: hard data_files: - split: train path: hard/train-* - config_name: medium data_files: - split: train path: medium/train-* --- # Adv_MATH500 Adversarial MATH500 prompts grouped by difficulty. ## Benchmark note When evaluated with `gpt5-nano`, `gpt5-nano` reaches: - `hard`: 34% accuracy - `medium`: 64% accuracy - `easy`: 93% accuracy