Base Fine Tuner

Bases: Bolt

An abstract base class for writing bolts for fine-tuning OpenAI models.

This base class is intended to be subclassed for fine-tuning OpenAI models. The chief objective of its subclasses is to load and preprocess the dataset, though of course, other methods, including fine-tuning, can be overridden for customization.

This bolt uses the OpenAI API to fine-tune a pre-trained model.

Each subclass can be invoked using the genius cli or yaml.

Using genius cli

genius <bolt_name> rise \
    batch \
        --input_s3_bucket my-input-bucket \
        --input_s3_folder my-input-folder \
    batch \
        --output_s3_bucket my-output-bucket \
        --output_s3_folder my-output-folder \
    postgres \
        --postgres_host 127.0.0.1 \
        --postgres_port 5432 \
        --postgres_user postgres \
        --postgres_password postgres \
        --postgres_database geniusrise \
        --postgres_table task_state \
    fine_tune \
    --args
        model=gpt-3.5-turbo \
        n_epochs=2 \
        batch_size=64 \
        learning_rate_multiplier=0.5 \
        prompt_loss_weight=1 \
        wait=True

This will load and preprocess data from input s3 location, and upload it to openai for fine tuning, and wait.

Using YAML

Bolts can be invoked using the genius cli on a yaml file.

Create a yaml file with the following content (looks very similar to cli):

version: 1

bolts:
    my_fine_tuner:
        name: OpenAIClassificationFineTuner
        method: fine_tune
        args:
            model: gpt-3.5-turbo
            n_epochs: 2
            batch_size: 64
            learning_rate_multiplier: 0.5
            prompt_loss_weight: 1
            wait: True
        input:
            type: batch
            bucket: my-input-bucket
            folder: my-input-folder
        output:
            type: batch
            bucket: my-output-bucket
            folder: my-output-folder
        state:
            type: postgres
            host: 127.0.0.1
            port: 5432
            user: postgres
            password: postgres
            database: geniusrise
            table: state

genius rise

Gotchas:

1. Extra command line arguments can be passed to the load_dataset method via fine_tune method by appending data_ to the param name.

e.g.

        args:
            model: gpt-3.5-turbo
            ...
            data_some_arbitrary_key: passed_to_load_dataset_method

__init__(input, output, state)

Initialize the bolt.

Parameters:

Name	Type	Description	Default
input	BatchInput	The batch input data.	required
output	BatchOutput	The output data.	required
state	State	The state manager.	required

delete_fine_tuned_model(model_id) staticmethod

Delete a fine-tuned model.

fine_tune(model, n_epochs, batch_size, learning_rate_multiplier, prompt_loss_weight, suffix=None, wait=False, data_extractor_lambda=None, **kwargs)

Fine-tune the model.

Parameters:

Name	Type	Description	Default
model	str	The pre-trained model name.	required
suffix	str	The suffix to append to the model name.	None
n_epochs	int	Total number of training epochs to perform.	required
batch_size	int	Batch size during training.	required
learning_rate_multiplier	int	Learning rate multiplier.	required
prompt_loss_weight	int	Prompt loss weight.	required
wait	bool	Whether to wait for the fine-tuning to complete. Defaults to False.	False
data_extractor_lambda	str	A lambda function run on each data element to extract the actual data.	None
**kwargs		Additional keyword arguments for training and data loading.	{}

Raises:

Type	Description
Exception	If any step in the fine-tuning process fails.

get_fine_tuning_job(job_id) staticmethod

Get the status of a fine-tuning job.

load_dataset(dataset_path, **kwargs) abstractmethod

Load a dataset from a file.

Parameters:

Name	Type	Description	Default
dataset_path	str	The path to the dataset file.	required
**kwargs		Additional keyword arguments to pass to the load_dataset method.	{}

Returns:

Name	Type	Description
Dataset	Union[Dataset, DatasetDict, Optional[Dataset]]	The loaded dataset.

Raises:

Type	Description
NotImplementedError	This method should be overridden by subclasses.

prepare_fine_tuning_data(data, data_type)

Prepare the given data for fine-tuning.

Parameters:

Name	Type	Description	Default
data	Union[Dataset, DatasetDict, Optional[Dataset]]	The dataset to prepare.	required
data_type	str	Either 'train' or 'eval' to specify the type of data.	required

Raises:

Type	Description
ValueError	If data_type is not 'train' or 'eval'.

preprocess_data(**kwargs)

Load and preprocess the dataset.

Raises:

Type	Description
Exception	If any step in the preprocessing fails.

wait_for_fine_tuning(job_id, check_interval=60)

Wait for a fine-tuning job to complete, checking the status every check_interval seconds.