nemo_rl.data.energon.multimodal.types#

Module Contents#

Classes#

MediaRef

One ordered media occurrence in a conversation.

CanonicalSFTSample

Model-neutral conversation produced by an Energon cooker.

EncodedSFTSample

One tokenized message log before batching.

PackedSFTSample

One physical pack of compatible encoded conversations.

Functions#

freeze_media_metadata

Validate and freeze media metadata without reading the media payload.

Data#

API#

nemo_rl.data.energon.multimodal.types.MediaMetadataValue: TypeAlias#

None

nemo_rl.data.energon.multimodal.types.FrozenMediaMetadata: TypeAlias#

None

nemo_rl.data.energon.multimodal.types.freeze_media_metadata(
metadata: object = None,
) nemo_rl.data.energon.multimodal.types.FrozenMediaMetadata#

Validate and freeze media metadata without reading the media payload.

class nemo_rl.data.energon.multimodal.types.MediaRef#

One ordered media occurrence in a conversation.

modality: str#

None

value: Any#

None

metadata: nemo_rl.data.energon.multimodal.types.FrozenMediaMetadata#

()

class nemo_rl.data.energon.multimodal.types.CanonicalSFTSample#

Bases: megatron.energon.Sample

Model-neutral conversation produced by an Energon cooker.

messages: list[dict[str, Any]]#

None

media: list[nemo_rl.data.energon.multimodal.types.MediaRef]#

None

tools: list[dict[str, Any]] | None#

None

class nemo_rl.data.energon.multimodal.types.EncodedSFTSample#

Bases: megatron.energon.Sample

One tokenized message log before batching.

message_log: list[dict[str, Any]]#

None

length: int#

None

packing_cost: int#

None

loss_multiplier: float#

None

group_key: tuple[Any, ...]#

None

sample_key: str#

None

pending_sample: nemo_rl.data.energon.multimodal.types.CanonicalSFTSample | None#

None

class nemo_rl.data.energon.multimodal.types.PackedSFTSample#

Bases: megatron.energon.Sample

One physical pack of compatible encoded conversations.

samples: list[nemo_rl.data.energon.multimodal.types.EncodedSFTSample]#

None

source_padded_lengths: list[int]#

None

group_key: tuple[Any, ...]#

None

pack_capacity: int#

None

nemo_rl.data.energon.multimodal.types.__all__#

[‘CanonicalSFTSample’, ‘EncodedSFTSample’, ‘FrozenMediaMetadata’, ‘MediaRef’, ‘MediaMetadataValue’, …