CalibrationConfig

The CalibrationConfig dataclass groups all calibration-related parameters used by Runner and AutoBitQuantizer.

CalibrationConfig dataclass

CalibrationConfig(calibration_dataset: str = 'c4', max_length: int = 2048, num_calibration_samples: int = 512, strategy: str = 'drop_rand', seed: int = 0, batch_size: Optional[int] = None, num_layers_per_group: int = 7, text_key: str = 'text', use_quality_filter: bool = False, max_documents: int = 10000)

Configuration for calibration data preparation.

Groups all calibration-related parameters into a single object used by :class:~onecomp.runner.Runner and :class:~onecomp.quantizer.autobit.AutoBitQuantizer.

Attributes:

Name	Type	Description
calibration_dataset	str	Dataset name ("c4", "wikitext2"), a local file path, or a HuggingFace Hub dataset ID. Defaults to "c4" (AllenAI C4 dataset).
max_length	int	Maximum token length per calibration chunk.
num_calibration_samples	int	Target number of calibration samples.
strategy	str	Chunking strategy. One of "concat_chunk", "concat_chunk_align", "concat_rand", "drop_head", "drop_rand".
seed	int	Random seed used by stochastic strategies (e.g. "drop_rand").
batch_size	int or None	Batch size for chunked calibration forward passes. None means a single forward pass with all data. Used only by chunked quantization (:func:~onecomp.runner_methods.chunked_quantization.run_chunked_quantization).
num_layers_per_group	int	Number of layers processed simultaneously in chunked calibration mode. Used only by chunked quantization (:func:~onecomp.runner_methods.chunked_quantization.run_chunked_quantization).
text_key	str	Column name used when loading custom or Hub datasets.
use_quality_filter	bool	Apply C4 quality filtering (ignored for non-C4 sources).
max_documents	int	Cap on documents loaded from custom files or Hub datasets.

Examples:

Default configuration (C4 dataset)::

config = CalibrationConfig()

WikiText-2 with longer sequences::

config = CalibrationConfig(
    calibration_dataset="wikitext2",
    max_length=2048,
    num_calibration_samples=256,
)

Custom file with quality filtering::

config = CalibrationConfig(
    calibration_dataset="./my_data.jsonl",
    text_key="content",
    use_quality_filter=True,
)