`nemo_rl.data`#

Subpackages#

Submodules#

Package Contents#

Classes#

`ResponseDatasetConfig`
`PreferenceDatasetConfig`
`DataConfig`
`MMLUEvalDataConfig`	Config for MMLU and multilingual MMLU datasets.
`MMLUProEvalDataConfig`	Config for MMLU Pro dataset.
`AIMEEvalDataConfig`	Config for AIME datasets loaded from the response registry.
`GPQAEvalDataConfig`	Config for GPQA datasets.
`MathEvalDataConfig`	Config for Math datasets.
`LocalMathEvalDataConfig`	Config for local math datasets loaded from files.
`MMAUEvalDataConfig`	Config for MMAU (Massive Multitask Audio Understanding) datasets.
`DailyOmniEvalDataConfig`	Config for the Daily-Omni audio-visual eval dataset.

Data#

EvalDataConfigType

API#

class nemo_rl.data.ResponseDatasetConfig[source]#

Bases: typing.TypedDict

dataset_name: NotRequired[str]#: None

data_path: NotRequired[str]#: None

input_key: NotRequired[str]#: None

output_key: NotRequired[str]#: None

subset: NotRequired[str | None]#: None

split: NotRequired[str]#: None

prompt_file: NotRequired[str | None]#: None

system_prompt_file: NotRequired[str | None]#: None

env_name: NotRequired[str]#: None

processor: NotRequired[str]#: None

download_dir: NotRequired[str]#: None

split_validation_size: NotRequired[float]#: None

seed: NotRequired[int]#: None

class nemo_rl.data.PreferenceDatasetConfig[source]#

Bases: typing.TypedDict

dataset_name: NotRequired[str]#: None

data_path: NotRequired[str]#: None

prompt_key: NotRequired[str]#: None

chosen_key: NotRequired[str]#: None

rejected_key: NotRequired[str]#: None

subset: NotRequired[str | None]#: None

split: NotRequired[str]#: None

prompt_file: NotRequired[str | None]#: None

system_prompt_file: NotRequired[str | None]#: None

class nemo_rl.data.DataConfig[source]#

Bases: typing.TypedDict

max_input_seq_length: int | None#: None

add_bos: NotRequired[bool]#: None

add_eos: NotRequired[bool]#: None

add_generation_prompt: NotRequired[bool]#: None

add_system_prompt: NotRequired[bool]#: None

shuffle: bool#: None

num_workers: NotRequired[int]#: None

use_multiple_dataloader: NotRequired[bool]#: None

num_prompts_per_dataloader: NotRequired[int]#: None

custom_dataloader: NotRequired[str]#: None

train: nemo_rl.data.ResponseDatasetConfig | nemo_rl.data.PreferenceDatasetConfig | list[nemo_rl.data.ResponseDatasetConfig]#: None

validation: NotRequired[nemo_rl.data.ResponseDatasetConfig | nemo_rl.data.PreferenceDatasetConfig | list[nemo_rl.data.ResponseDatasetConfig] | None]#: None

default: NotRequired[nemo_rl.data.ResponseDatasetConfig | nemo_rl.data.PreferenceDatasetConfig | None]#: None

class nemo_rl.data.MMLUEvalDataConfig[source]#

Bases: typing.TypedDict

Config for MMLU and multilingual MMLU datasets.

Supports dataset_name: “mmlu” or “mmlu_{language}” where language is one of: AR-XY, BN-BD, DE-DE, EN-US, ES-LA, FR-FR, HI-IN, ID-ID, IT-IT, JA-JP, KO-KR, PT-BR, ZH-CN, SW-KE, YO-NG

Initialization

Initialize self. See help(type(self)) for accurate signature.

max_input_seq_length: int#: None

dataset_name: Literal[mmlu, mmlu_AR-XY, mmlu_BN-BD, mmlu_DE-DE, mmlu_EN-US, mmlu_ES-LA, mmlu_FR-FR, mmlu_HI-IN, mmlu_ID-ID, mmlu_IT-IT, mmlu_JA-JP, mmlu_KO-KR, mmlu_PT-BR, mmlu_ZH-CN, mmlu_SW-KE, mmlu_YO-NG]#: None

prompt_file: NotRequired[str | None]#: None

system_prompt_file: NotRequired[str | None]#: None

class nemo_rl.data.MMLUProEvalDataConfig[source]#

Bases: typing.TypedDict

Config for MMLU Pro dataset.

Initialization

Initialize self. See help(type(self)) for accurate signature.

max_input_seq_length: int#: None

dataset_name: Literal[mmlu_pro]#: None

prompt_file: NotRequired[str | None]#: None

system_prompt_file: NotRequired[str | None]#: None

class nemo_rl.data.AIMEEvalDataConfig[source]#

Bases: typing.TypedDict

Config for AIME datasets loaded from the response registry.

Initialization

Initialize self. See help(type(self)) for accurate signature.

max_input_seq_length: int#: None

dataset_name: Literal[AIME2024, AIME2025, AIME2026]#: None

prompt_file: NotRequired[str | None]#: None

system_prompt_file: NotRequired[str | None]#: None

processor: NotRequired[str]#: None

repeat: NotRequired[int]#: None

class nemo_rl.data.GPQAEvalDataConfig[source]#

Bases: typing.TypedDict

Config for GPQA datasets.

Initialization

Initialize self. See help(type(self)) for accurate signature.

max_input_seq_length: int#: None

dataset_name: Literal[gpqa, gpqa_diamond]#: None

prompt_file: NotRequired[str | None]#: None

system_prompt_file: NotRequired[str | None]#: None

class nemo_rl.data.MathEvalDataConfig[source]#

Bases: typing.TypedDict

Config for Math datasets.

Initialization

Initialize self. See help(type(self)) for accurate signature.

max_input_seq_length: int#: None

dataset_name: Literal[math, math500]#: None

prompt_file: NotRequired[str | None]#: None

system_prompt_file: NotRequired[str | None]#: None

class nemo_rl.data.LocalMathEvalDataConfig[source]#

Bases: typing.TypedDict

Config for local math datasets loaded from files.

dataset_name can be a URL or local file path. Requires additional fields: problem_key, solution_key, file_format, split.

Initialization

Initialize self. See help(type(self)) for accurate signature.

max_input_seq_length: int#: None

dataset_name: str#: None

problem_key: str#: None

solution_key: str#: None

file_format: Literal[csv, json]#: None

split: NotRequired[str | None]#: None

prompt_file: NotRequired[str | None]#: None

system_prompt_file: NotRequired[str | None]#: None

class nemo_rl.data.MMAUEvalDataConfig[source]#

Bases: typing.TypedDict

Config for MMAU (Massive Multitask Audio Understanding) datasets.

Initialization

Initialize self. See help(type(self)) for accurate signature.

max_input_seq_length: int#: None

dataset_name: Literal[mmau, TwinkStart/MMAU]#: None

split: NotRequired[str | None]#: None

prompt_file: NotRequired[str | None]#: None

system_prompt_file: NotRequired[str | None]#: None

env_name: NotRequired[str]#: None

class nemo_rl.data.DailyOmniEvalDataConfig[source]#

Bases: typing.TypedDict

Config for the Daily-Omni audio-visual eval dataset.

Mirrors the MMAU multimodal schema but with its own dataset_name literal so the eval-config union resolves daily-omni unambiguously. Kept as a TypedDict for consistency with the other (still v1) eval-data configs in this union, whose consumers access the resolved config by key (config.data["dataset_name"]).

Fields: max_input_seq_length: Max prompt length passed to the generation backend. dataset_name: Must be "daily-omni". split: HuggingFace split to load. prompt_file: Optional prompt template path. system_prompt_file: Optional system prompt path. env_name: Reward/eval environment name (e.g. "vlm").

Initialization

Initialize self. See help(type(self)) for accurate signature.

max_input_seq_length: int#: None

dataset_name: Literal[daily-omni]#: None

split: NotRequired[str | None]#: None

prompt_file: NotRequired[str | None]#: None

system_prompt_file: NotRequired[str | None]#: None

env_name: NotRequired[str]#: None

nemo_rl.data.EvalDataConfigType#: None

nemo_rl.data#

Subpackages#

Submodules#

Package Contents#

Classes#

Data#

API#

`nemo_rl.data`#