bridge.data.builders.mock_vlm_sft

Serializable config and runtime builder for synthetic VLM conversations.

Module Contents

Classes

MockVLMSFTDatasetConfig	Serializable settings for synthetic conversation-style VLM data.
MockVLMSFTDatasetBuilder	Build synthetic VLM datasets from declarative settings.

Functions

make_mock_vlm_example	Create one synthetic conversation with the configured number of images.
make_mock_vlm_examples	Generate the deterministic base examples reused by every requested split.
build_mock_vlm_sft_split	Build one requested synthetic VLM split.
mock_vlm_sft_train_valid_test_datasets_provider	Build synthetic VLM splits through the canonical runtime builder.

Data

_MOCK_RESPONSE_VOCABULARY

API

bridge.data.builders.mock_vlm_sft._MOCK_RESPONSE_VOCABULARY

‘split(…)’

class bridge.data.builders.mock_vlm_sft.MockVLMSFTDatasetConfig

Bases: megatron.bridge.data.base.DataloaderConfig

Serializable settings for synthetic conversation-style VLM data.

seq_length: int

None

hf_processor_path: str

None

prompt: str

‘Describe this image.’

random_seed: int

0

image_size: tuple[int, int]

(256, 256)

num_images: int

1

num_base_examples: int

1000

skip_getting_attention_mask_from_dataset: bool

True

dataloader_type: Literal[single, cyclic] | None

‘single’

enable_in_batch_packing: bool

False

defer_in_batch_packing_to_step: bool

False

pad_to_max_length: bool

False

pad_to_multiple_of: int

128

in_batch_packing_pad_to_multiple_of: int

1

validate() → None

Validate synthetic data settings.

finalize() → None

Finalize dataloader settings and validate this config.

bridge.data.builders.mock_vlm_sft.make_mock_vlm_example(

config: bridge.data.builders.mock_vlm_sft.MockVLMSFTDatasetConfig,

rng: numpy.random.Generator,

response_text: str,

) → dict[str, Any]

Create one synthetic conversation with the configured number of images.

bridge.data.builders.mock_vlm_sft.make_mock_vlm_examples(

config: bridge.data.builders.mock_vlm_sft.MockVLMSFTDatasetConfig,

) → list[dict[str, Any]]

Generate the deterministic base examples reused by every requested split.

bridge.data.builders.mock_vlm_sft.build_mock_vlm_sft_split(

config: bridge.data.builders.mock_vlm_sft.MockVLMSFTDatasetConfig,

base_examples: list[dict[str, Any]],

target_length: int,

processor: Any,

) → megatron.bridge.data.datasets.direct_sft.DirectSFTDataset | None

Build one requested synthetic VLM split.

class bridge.data.builders.mock_vlm_sft.MockVLMSFTDatasetBuilder(

config: bridge.data.builders.mock_vlm_sft.MockVLMSFTDatasetConfig,

)

Build synthetic VLM datasets from declarative settings.

Initialization

build(

context: megatron.bridge.data.base.DatasetBuildContext,

) → tuple[megatron.bridge.data.datasets.direct_sft.DirectSFTDataset | None, megatron.bridge.data.datasets.direct_sft.DirectSFTDataset | None, megatron.bridge.data.datasets.direct_sft.DirectSFTDataset | None]

Build the train, validation, and test splits requested by the schedule.

bridge.data.builders.mock_vlm_sft.mock_vlm_sft_train_valid_test_datasets_provider(

train_val_test_num_samples: list[int],

dataset_config: bridge.data.builders.mock_vlm_sft.MockVLMSFTDatasetConfig,

tokenizer: Any | None = None,

pg_collection: Any | None = None,

) → tuple[megatron.bridge.data.datasets.direct_sft.DirectSFTDataset | None, megatron.bridge.data.datasets.direct_sft.DirectSFTDataset | None, megatron.bridge.data.datasets.direct_sft.DirectSFTDataset | None]

Build synthetic VLM splits through the canonical runtime builder.