Supervised Fine-Tuning with Hugging Face
Learn about the hf_sft function within Simplifine’s Train Engine
hf_sft(
model_name:str,
dataset_name:str='nlpie/pandemic_pact',
keys:list=[],
template:str='',
do_split:bool=True,
split_ratio:float=0.2,
load_eval_from_data:bool=False,
data:dict={},
num_epochs:int=3,
batch_size:int=1,
wandb_api_key:str='',
lr:float=5e-5,
from_hf:bool=True,
response_template:str='### Answer:',
eval_steps:int=10,
logging_steps:int=10,
use_peft:bool=False,
peft_config=None,
ddp:bool=False,
zero:bool=True,
deepspeed_config:str='home/ubuntu/src/zero_config.json',
hf_token:str='',
gradient_accumulation_steps:int=1,
fp16:bool=False,
bf16:bool=False,
report_to:str='none',
gradient_checkpointing:bool=False,
max_seq_length:int=2048,
use_wandb:bool=False,
output_dir:str='sft_output',
eval_accumulation_steps:int=8,
wandb_config:wandbConfig=None
):
The name or path of the pre-trained model to use.
The name of the dataset to use for training. Defaults to 'nlpie/pandemic_pact'.
List of keys used for formatting prompts in the template. Defaults to an empty list.
Template string for formatting prompts with keys. Defaults to an empty string.
A flag to determine whether to split the dataset into training and validation sets. Defaults to True.
The ratio of the dataset to be used for validation. Defaults to 0.2.
A flag to determine whether to load evaluation data from the provided dataset. Defaults to False.
A dictionary containing the training data. Defaults to an empty dictionary.
The number of training epochs. Defaults to 3.
The batch size for training. Defaults to 1.
The API key for Weights and Biases (WandB) logging. Defaults to an empty string.
The learning rate for optimization. Defaults to 5e-5.
A flag to determine whether to load the dataset from Hugging Face. Defaults to True.
Template string for response formatting. Defaults to '### Answer:'.
The number of steps between evaluations. Defaults to 10.
The number of steps between logging outputs. Defaults to 10.
A flag to enable Parameter-Efficient Fine-Tuning (PEFT). Defaults to False.
The configuration object for PEFT. Defaults to None.
A flag to enable Distributed Data Parallel (DDP) training. Defaults to False.
A flag to enable ZeRO (Zero Redundancy Optimizer) for memory optimization. Defaults to True.
The path to the DeepSpeed configuration file. Defaults to 'home/ubuntu/src/zero_config.json'.
The Hugging Face token required for accessing private datasets or models. Defaults to an empty string.
The number of steps for gradient accumulation. Defaults to 1.
A flag to enable 16-bit floating-point (FP16) training. Defaults to False.
A flag to enable 16-bit Brain Floating Point (BF16) training. Defaults to False.
The service to report training logs to (e.g., wandb). Defaults to 'none'.
A flag to enable gradient checkpointing for reducing memory usage. Defaults to False.
The maximum sequence length for the input data. Defaults to 2048.
A flag to enable Weights and Biases (WandB) logging. Defaults to False.
The directory to save the output model and logs. Defaults to 'sft_output'.
The number of steps for evaluation accumulation. Defaults to 8.
The configuration for Weights and Biases (WandB) logging. Defaults to None.