adaptation

adaptation

Configuration classes for model adaptation for parameters-efficient fine-tuning.

LoRAAdaptationConfig dataclass

LoRAAdaptationConfig(_target_='versatil.models.adaptation.lora.LoRAAdaptation', enabled=True, rank=8, alpha=16, dropout=0.0, target_modules=value, exclude_modules=None, bias='none', init_lora_weights=DEFAULT_LORA_INIT_WEIGHTS)

Hydra configuration for Low-Rank Adaptation.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
enabled	bool	Whether Hydra should instantiate LoRA adaptation for the target module.
rank	int	Adapter rank r. Higher ranks increase adapter capacity and trainable parameters.
alpha	int	Adapter scaling factor. PEFT applies the learned update with scale alpha / rank.
dropout	float	Dropout probability on the adapter path before the low-rank update.
target_modules	str	PEFT target-module preset.
exclude_modules	list[str] | None	Optional module names to leave unwrapped by LoRA.
bias	str	PEFT bias handling mode.
init_lora_weights	str	PEFT adapter weight initialization strategy.