easydel.trainers.supervised_fine_tuning_trainer.__init__

class easydel.trainers.supervised_fine_tuning_trainer.__init__.SFTConfig(auto_shard_states: bool = True, aux_loss_enabled: bool = False, backend: tp.Optional[str] = None, clip_grad: tp.Optional[float] = None, custom_scheduler: tp.Optional[tp.Callable[[int], tp.Any]] = None, dataloader_num_workers: tp.Optional[int] = 0, dataloader_pin_memory: tp.Optional[bool] = False, do_eval: bool = False, do_last_save: bool = True, do_train: bool = True, eval_batch_size: tp.Optional[int] = None, evaluation_steps: tp.Optional[int] = None, extra_optimizer_kwargs: dict = <factory>, frozen_parameters: tp.Optional[str] = None, gradient_accumulation_steps: int = 1, ids_to_pop_from_dataset: tp.Optional[tp.List[str]] = <factory>, is_fine_tuning: bool = True, init_tx: bool = True, jax_distributed_config: tp.Optional[dict] = None, learning_rate: float = 2e-05, learning_rate_end: tp.Optional[float] = None, log_all_workers: bool = False, log_grad_norms: bool = True, report_metrics: bool = True, log_steps: int = 10, loss_config: tp.Optional[LossConfig] = None, low_mem_usage: bool = True, max_evaluation_steps: tp.Optional[int] = None, max_sequence_length: tp.Optional[int] = 4096, max_training_steps: tp.Optional[int] = None, model_name: str = 'SFTTrainer', model_parameters: tp.Optional[dict] = None, metrics_to_show_in_rich_pbar: tp.Optional[tp.List[str]] = None, num_train_epochs: int = 10, offload_dataset: bool = False, offload_device_type: str = 'cpu', offload_device_index: int = 0, optimizer: AVAILABLE_OPTIMIZERS = EasyDeLOptimizers.ADAMW, performance_mode: bool = False, pruning_module: tp.Any = None, process_zero_is_admin: bool = True, progress_bar_type: tp.Literal['tqdm', 'rich', 'json'] = 'tqdm', remove_ckpt_after_load: bool = False, remove_unused_columns: bool = True, report_steps: int = 5, save_directory: str = 'EasyDeL-Checkpoints', save_optimizer_state: bool = True, save_steps: tp.Optional[int] = None, save_total_limit: tp.Optional[int] = None, scheduler: AVAILABLE_SCHEDULERS = EasyDeLSchedulers.NONE, sparsify_module: bool = False, sparse_module_type: AVAILABLE_SPARSE_MODULE_TYPES = 'bcoo', state_apply_fn_kwarguments_to_model: tp.Optional[dict] = None, step_partition_spec: PartitionSpec = PartitionSpec(('dp', 'fsdp'), 'sp'), step_start_point: tp.Optional[int] = None, shuffle_train_dataset: bool = True, total_batch_size: int = 32, training_time_limit: tp.Optional[str] = None, train_on_inputs: bool = True, truncation_mode: tp.Literal['keep_end', 'keep_start'] = 'keep_end', tx_mu_dtype: tp.Optional[jnp.dtype] = None, track_memory: bool = False, use_data_collactor: bool = True, use_wandb: bool = True, verbose: bool = True, wandb_entity: tp.Optional[str] = None, warmup_steps: int = 0, weight_decay: float = 0.01, dataset_text_field: ~typing.Optional[str] = None, add_special_tokens: bool = False, packing: bool = False, dataset_num_proc: ~typing.Optional[int] = None, dataset_batch_size: int = 1000, dataset_kwargs: ~typing.Optional[dict[str, typing.Any]] = None, eval_packing: ~typing.Optional[bool] = None, num_of_sequences: int = 1024, chars_per_token: float = 3.6)

Bases: TrainingArguments

Configuration class for the [SFTTrainer].

Parameters

• model_name (str) – The name of the model. Defaults to “SFTTrainer”.

• dataset_text_field (str, optional) – Name of the text field of the dataset. If provided, the trainer will automatically create a [ConstantLengthDataset] based on dataset_text_field. Defaults to None.

• packing (bool, optional) – Controls whether the [ConstantLengthDataset] packs the sequences of the dataset. Defaults to False.

• learning_rate (float, optional) – Initial learning rate for [AdamW] optimizer. The default value replaces that of [~transformers.TrainingArguments]. Defaults to 2e-5.

• dataset_num_proc (int, optional) – Number of processes to use for processing the dataset. Only used when packing=False. Defaults to None.

• dataset_batch_size (int, optional) – Number of examples to tokenize per batch. If dataset_batch_size <= 0 or dataset_batch_size is None, tokenizes the full dataset as a single batch. Defaults to 1000.

• dataset_kwargs (dict[str, Any], optional) – Dictionary of optional keyword arguments to pass when creating packed or non-packed datasets. Defaults to None.

• eval_packing (bool, optional) – Whether to pack the eval dataset. If None, uses the same value as packing. Defaults to None.

• num_of_sequences (int, optional) – Number of sequences to use for the [ConstantLengthDataset]. Defaults to 1024.

• chars_per_token (float, optional) – Number of characters per token to use for the [ConstantLengthDataset]. See [chars_token_ratio](huggingface/trl) for more details. Defaults to 3.6.

add_special_tokens: bool = False

chars_per_token: float = 3.6

dataset_batch_size: int = 1000

dataset_kwargs: Optional[dict[str, Any]] = None

dataset_num_proc: Optional[int] = None

dataset_text_field: Optional[str] = None

eval_packing: Optional[bool] = None

extra_optimizer_kwargs: dict

ids_to_pop_from_dataset: tp.Optional[tp.List[str]]

learning_rate: float = 2e-05

model_name: str = 'SFTTrainer'

num_of_sequences: int = 1024

packing: bool = False

class easydel.trainers.supervised_fine_tuning_trainer.__init__.SFTTrainer(arguments: SFTConfig, processing_class: Any, model: Optional[Union[EasyDeLBaseModule, EasyDeLState]] = None, train_dataset: Optional[Any] = None, eval_dataset: Optional[Union[Any, Dict[str, Any]]] = None, formatting_func: Optional[Callable] = None, data_collator: Optional[DataCollatorForCompletionOnlyLM] = None)

Bases: Trainer

Trainer class for Supervised Fine-Tuning (SFT) of language models.

This trainer extends the Trainer and provides functionalities specific to supervised fine-tuning tasks.