Ft libraries

Base interfaces for fine-tuning framework libraries.

This module defines the protocols and base classes for implementing fine-tuning libraries that can be used with the gllm-training library.

BaseFineTuningLibrary

Bases: ABC

Base class for fine-tuning library implementations that extends Component.

framework_name: str property

Return the name of the framework.

Returns:

Type	Description
str	Framework name (e.g., "unsloth", "transformers").

load_model(experiment_args, hyperparameters, **kwargs) abstractmethod

Load a model based on the provided configuration.

Parameters:

Name	Type	Description	Default
experiment_args	ExperimentConfig	Configuration for the experiment.	required
hyperparameters	FinetunedHyperparameters	Configuration for the hyperparameters.	required
**kwargs	Any	Additional framework-specific arguments.	{}

Returns:

Type	Description
	A tuple containing the loaded model and tokenizer.

save_model(model, path, **kwargs) abstractmethod

Save the model to the specified path.

The associated tokenizer should also be saved, often to the same path.

Parameters:

Name	Type	Description	Default
model		The model instance to save.	required
path	str	The directory path to save the model to.	required
**kwargs	Any	Additional framework-specific arguments for saving.	{}

save_tokenizer(tokenizer, path, **kwargs) abstractmethod

Saves the tokenizer to the specified path.

Parameters:

Name	Type	Description	Default
tokenizer		The tokenizer instance to save.	required
path	str	The directory path to save the tokenizer to.	required
**kwargs	Any	Additional framework-specific arguments.	{}

setup_training(model, tokenizer, hyperparameters, train_dataset, eval_dataset=None, dataset_text_field='prompt', **kwargs) abstractmethod

Set up a trainer for the model based on the provided configuration.

Parameters:

Name	Type	Description	Default
model		The model instance to train.	required
tokenizer		The tokenizer instance to use with the model and data.	required
hyperparameters	FinetunedHyperparameters	Configuration for the hyperparameters.	required
train_dataset	Dataset	The training dataset.	required
eval_dataset	Dataset	The evaluation dataset (optional).	None
dataset_text_field	str	Field in dataset containing the text data. Defaults to "prompt".	'prompt'
**kwargs	Any	Additional framework-specific arguments.	{}

Returns:

Type	Description
	The initialized trainer instance.

FineTuningLibraryProtocol

Bases: Protocol[ModelType_co, TrainerType_co, TokenizerType]

Protocol that all fine-tuning frameworks must implement.

framework_name: str property

Return the name of the framework.

Returns:

Type	Description
str	Framework name (e.g., "unsloth", "transformers").

load_model(experiment_args, hyperparameters, **kwargs)

Load a model based on the provided configuration.

The associated tokenizer should also be loaded/initialized internally and be retrievable via get_tokenizer().

Parameters:

Name	Type	Description	Default
experiment_args	ExperimentConfig	Configuration for the experiment.	required
hyperparameters	FinetunedHyperparameters	Configuration for the hyperparameters.	required
**kwargs	Any	Additional framework-specific arguments.	{}

Returns:

Type	Description
Tuple[ModelType_co, TokenizerType]	A tuple containing the loaded model and tokenizer.

save_model(model, path, **kwargs)

Save the model to the specified path.

save_tokenizer(tokenizer, path, **kwargs)

Saves the tokenizer to the specified path.

setup_training(model, tokenizer, hyperparameters, train_dataset, eval_dataset=None, dataset_text_field='prompt', **kwargs)

Set up a trainer for the model based on the provided configuration.