Finetune

Dataset

Dataset(path: Path | str | None)

Dataset to finetune LLMs on generating synthetic data.

Parameters:

Name	Type	Description	Default
path	Path | str | None	The path to the dataset on disk. If None, the dataset will be stored to a file in a newly created temporary directory.	required

append

append(task: BaseTask, data: RelationalData) -> None

Append data to the dataset.

Parameters:

Name	Type	Description	Default
task	BaseTask	The task to append the data for.	required
data	RelationalData	The data to be appended.	required

load

load() -> Dataset

Load the dataset as a Hugging Face datasets.Dataset.

Returns:

Type	Description
Dataset	The Hugging Face datasets.Dataset.

load_prompt_completion

load_prompt_completion(
    prompt_template: str | Callable[[DatasetElem], str],
    eos_token: str = "",
    tokenizer: PreTrainedTokenizer | None = None,
    tokenize: bool = False,
    completion_key: str = "out",
    max_length: int | None = None,
    enforce_max_len: bool = True,
) -> Dataset

Load the dataset as a Hugging Face dataset, map it to prompt-completion format and (optionally) tokenize it.

Parameters:

Name	Type	Description	Default
prompt_template	str | Callable[[DatasetElem], str]	Either a string to be formatted to get the prompts, or a callable returning the prompt from the JSON representation of DatasetElem. If a string, it may contain as keys the fields of DatasetElem.	required
eos_token	str	An optional EOS token to add to the end of the completion. Should not be given if the tokenizer is provided.	''
tokenizer	PreTrainedTokenizer | None	A Hugging Face tokenizer. If given, its EOS is added at the end of the completion, and it is used to optionally tokenize the dataset.	None
tokenize	bool	Whether to tokenize the dataset. If True, the trainer will skip the tokenization step.	False
completion_key	str	The field of DatasetElem to be used as completion.	'out'
max_length	int | None	The maximum length in tokens of a single example.	None
enforce_max_len	bool	Whether to raise an error in case the number of tokens of a single example exceeds max_length. If False, a simple warning is issued.	True

Returns:

Type	Description
Dataset	The Hugging Face datasets.Dataset in prompt-completion form.

FtConfig pydantic-config

FtConfig(
    dataset_path: str | Path | None,
    prompt_template: str | Callable[[DatasetElem], str],
    enforce_max_len: bool = True,
    valid_frac: Annotated[float, Field(gt=0, lt=1)] = 0.1,
    early_stop: NonNegativeInt = 0,
    dump_n_examples: NonNegativeInt = 10,
    best_ckpt: str = "best",
)

Finetune configuration.

Parameters:

Name	Type	Description	Default
dataset_path	str | Path | None	The path to the dataset. If the dataset has to be created and the path is set to None, the dataset is saved in a temporary directory.	required
prompt_template	str | Callable[[DatasetElem], str]	The template for the prompt. It may contain as keys the fields of DatasetElem.	required
enforce_max_len	bool	Whether to enforce the max length coming from the model during dataset loading (in the prompt-completion format).	True
valid_frac	Annotated[float, Field(gt=0, lt=1)]	Fraction of the data to be used for validation.	0.1
early_stop	NonNegativeInt	Early stopping patience.	0
dump_n_examples	NonNegativeInt	Number of dataset examples to dump for inspection.	10
best_ckpt	str	Name of the best checkpoint.	'best'

finetune

finetune

finetune(
    cfg_ft: FtConfig,
    cfg_model: ModelConfig,
    cfg_sft: SFTConfig,
    unsloth: bool = False,
) -> None

Finetune an LLM using a given Dataset.

Parameters:

Name	Type	Description	Default
cfg_ft	FtConfig	The finetune configuration.	required
cfg_model	ModelConfig	The model configuration.	required
cfg_sft	SFTConfig	The trainer configuration.	required
unsloth	bool	Whether to use unsloth to perform the finetuning. If not, the standard Hugging Face finetune will be used.	False