nemo_rl.data.datasets.response_datasets.tulu3#

Module Contents#

Classes#

Tulu3SftMixtureDataset

Tulu3 SFT mixture dataset.

Functions#

format_tulu3_sft_mixture

Format for Tulu3 SFT data.

API#

nemo_rl.data.datasets.response_datasets.tulu3.format_tulu3_sft_mixture(
data: dict[str, Any],
task_name: str = 'tulu3_sft_mixture',
) → dict[str, str | dict[str, str]]#

Format for Tulu3 SFT data.

class nemo_rl.data.datasets.response_datasets.tulu3.Tulu3SftMixtureDataset(
seed: int = 42,
test_size: float = 0.05,
prompt_file: str | None = None,
max_samples: int | None = None,
)#

Bases: nemo_rl.data.datasets.raw_dataset.RawDataset

Tulu3 SFT mixture dataset.

Initialization

Initialize the Tulu3 SFT mixture dataset.

Parameters:

  • seed – Random seed for train/validation split

  • test_size – Proportion of data to use for validation (0.0-1.0)

  • prompt_file – Optional prompt file path to be applied via TaskDataSpec

  • max_samples – Optional maximum number of samples to use from the dataset