sagemaker.train.common

Classes

CustomizationTechnique(value)	Customization techniques for fine-tuning.
FineTuningOptions(options_dict[, ...])	Dynamic class for fine-tuning options with validation.
TrainingType(value)	Training types for fine-tuning.

class sagemaker.train.common.CustomizationTechnique(value)

Bases: Enum

Customization techniques for fine-tuning.

CPT = 'CPT'

DPO = 'DPO'

RLAIF = 'RLAIF'

RLVR = 'RLVR'

SFT = 'SFT'

class sagemaker.train.common.FineTuningOptions(options_dict: Dict[str, Any], sequence_length: int | None = None)

Bases: object

Dynamic class for fine-tuning options with validation.

get_info(param_name: str = None)

Display parameter information in a user-friendly format.

to_dict() → Dict[str, Any]

Convert back to dictionary for hyperparameters with string values.

to_user_dict() → Dict[str, Any]

Return only user-explicitly-set hyperparameters as string key-value pairs.

validate_length_constraints()

Enforce that selected lengths fit the recipe’s sequence_length.

For RL recipes: max_prompt_length + max_response_length must not exceed sequence_length. For SFT/DPO: the single-example length (dataset_max_len) must not exceed it. Per-field min/max are already enforced on assignment; this adds the cross-field sum check that a per-field max cannot express.

Framework-agnostic: gated only on the recipe’s sequence_length metadata, not on the model family. No-op if sequence_length is unknown or the relevant params are absent.

class sagemaker.train.common.TrainingType(value)

Bases: Enum

Training types for fine-tuning.

FULL = 'FULL'

LORA = 'LORA'