bridge.data.vlm_datasets.mock_provider

Generic mock conversation-style VLM dataset and provider.

This module produces synthetic image(s) and minimal conversations that are compatible with HF AutoProcessor.apply_chat_template and the collate functions defined in collate.py. It is processor-agnostic and can be used with any multimodal model whose processor supports the standard conversation schema and optional images argument.

Module Contents

Classes

MockVLMConversationProvider

DatasetProvider for generic mock VLM conversation datasets.

API

class bridge.data.vlm_datasets.mock_provider.MockVLMConversationProvider

Bases: megatron.bridge.training.config.DatasetProvider

DatasetProvider for generic mock VLM conversation datasets.

Builds train/valid/test datasets using a HF AutoProcessor and the MockVLMConversationDataset implementation. Intended to work across different VLM models whose processors support the conversation schema.

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)

pad_to_max_length: bool

True

create_attention_mask: bool

True

skip_getting_attention_mask_from_dataset: bool

True

num_images: int

1

dataloader_type: Optional[Literal[single, cyclic, external]]

‘single’

_processor: Optional[Any]

None

_make_base_examples() → List[Dict[str, Any]]

build_datasets(

context: megatron.bridge.training.config.DatasetBuildContext,

)