`nemo_rl.data.datasets.response_datasets.mmpr_tiny`#

Module Contents#

Classes#

MMPRTinyDataset

Wrapper around the MMPR-Tiny dataset (OpenGVLab/MMPR-Tiny).

Functions#

`format_mmpr_tiny_dataset`	Format the MMPR-Tiny dataset into an OpenAI-API-like message log.
`_ensure_mmpr_cached`	Download and extract MMPR-Tiny images if not already cached.
`_load_mmpr_tiny_from_cache`	Load MMPR-Tiny dataset from a preprocessed cache directory.

Data#

_IMAGE_PLACEHOLDER_RE

API#

nemo_rl.data.datasets.response_datasets.mmpr_tiny._IMAGE_PLACEHOLDER_RE#: ‘compile(…)’

nemo_rl.data.datasets.response_datasets.mmpr_tiny.format_mmpr_tiny_dataset( example: dict[str, Any], ) → dict[str, Any]#

Format the MMPR-Tiny dataset into an OpenAI-API-like message log.

Supports multi-image rows by splitting question text on and <image_N> placeholders and interleaving image content items with text segments. Each placeholder consumes the next image path in order.

nemo_rl.data.datasets.response_datasets.mmpr_tiny._ensure_mmpr_cached(download_dir: str) → None#

Download and extract MMPR-Tiny images if not already cached.

Thread-safe: uses atomic marker file to prevent race conditions in distributed settings.

nemo_rl.data.datasets.response_datasets.mmpr_tiny._load_mmpr_tiny_from_cache(download_dir: str) → datasets.Dataset#: Load MMPR-Tiny dataset from a preprocessed cache directory.

class nemo_rl.data.datasets.response_datasets.mmpr_tiny.MMPRTinyDataset(

split: str = 'train',

download_dir: str = '',

split_validation_size: float = 0,

seed: int = 42,

**kwargs,

)#

Bases: nemo_rl.data.datasets.raw_dataset.RawDataset

Wrapper around the MMPR-Tiny dataset (OpenGVLab/MMPR-Tiny).

Supports loading from a local preprocessed cache or downloading from HuggingFace.

Parameters:

split – Dataset split to use. Only “train” is supported (validation is created via split_validation_size).
download_dir – Directory containing the preprocessed MMPR-Tiny cache, or the target directory for downloading from HuggingFace.
split_validation_size – Fraction of data to hold out for validation (default 0).
seed – Random seed for train/validation split (default 42).

Initialization

nemo_rl.data.datasets.response_datasets.mmpr_tiny#