nemo_rl.data.datasets.response_datasets.aime#

AIME response dataset (2024/2025/2026 variants).

Module Contents#

Classes#

AIMEDataset

Wrapper around an AIME competition dataset with a train split.

Data#

AIMEVariant

API#

nemo_rl.data.datasets.response_datasets.aime.AIMEVariant#

None

class nemo_rl.data.datasets.response_datasets.aime.AIMEDataset(
variant: nemo_rl.data.datasets.response_datasets.aime.AIMEVariant = '2024',
repeat: int = 1,
**kwargs,
)#

Bases: nemo_rl.data.datasets.raw_dataset.RawDataset

Wrapper around an AIME competition dataset with a train split.

Parameters:

  • variant – Which AIME edition to load: “2024” (default), “2025”, or “2026”.

  • repeat – Number of times to repeat the dataset, default is 1. Training recipes that need repeated validation should set this explicitly.

Initialization

format_data(data: dict[str, Any]) → dict[str, Any]#