Class TrainingArguments

main_e2e_qp.py:136–194

@dataclass
class TrainingArguments(transformers.Seq2SeqTrainingArguments):
    cache_dir: Optional[str] = field(
        default=None
    )
    train_on_source: Optional[bool] = field(
        default=False,
        metadata={"help": "Whether to train on the input in addition to the target text."}
    )
    do_mmlu_eval: Optional[bool] = field(
        default=False,
        metadata={"help": "Whether to run the MMLU evaluation."}
    )
    do_ppl_eval: Optional[bool] = field(
        default=False,
        metadata={"help": "Whether to run the PPL evaluation."}
    )
    pt_context_len: int = field(
        default=1024,
        metadata={"help": "language modeling length."}
    )
    full_finetune: bool = field(
        default=False,
        metadata={"help": "Finetune the entire model without adapters."}
    )
    wbits: int = field(
        default=4,
        metadata={"help": "How many bits to use."}
    )
    group_size: int = field(
        default=64,
        metadata={"help": "How many group size to use."}
    )
    max_memory_MB: int = field(
        default=80000,
        metadata={"help": "Free memory per gpu."}
    )
    report_to: str = field(
        default='none',
        metadata={"help": "To use wandb or something else for reporting."}
    )
    output_dir: str = field(default='./output', metadata={"help": 'The output dir for logs and checkpoints'})
    resume_from_checkpoint: str = field(default=None, metadata={"help": 'The output dir for logs and checkpoints'})
    optim: str = field(default='paged_adamw_32bit', metadata={"help": 'The optimizer to be used'})
    per_device_train_batch_size: int = field(default=1, metadata={"help": 'The training batch size per GPU. Increase for better speed.'})
    gradient_accumulation_steps: int = field(default=16, metadata={"help": 'How many gradients to accumulate before to perform an optimizer step'})
    max_steps: int = field(default=0, metadata={"help": 'How many optimizer update steps to take'})
    weight_decay: float = field(default=0.0, metadata={"help": 'The L2 weight decay rate of AdamW'}) # use lora dropout instead for regularization if needed
    learning_rate: float = field(default=2e-5, metadata={"help": 'The learnign rate'})
    remove_unused_columns: bool = field(default=False, metadata={"help": 'Removed unused columns. Needed to make this codebase work.'})
    max_grad_norm: float = field(default=0.3, metadata={"help": 'Gradient clipping max norm. This is tuned and works well for all models tested.'})
    gradient_checkpointing: bool = field(default=True, metadata={"help": 'Use gradient checkpointing. You want to use this.'})
    do_train: bool = field(default=True, metadata={"help": 'To train or not to train, that is the question?'})
    lr_scheduler_type: str = field(default='cosine', metadata={"help": 'Learning rate schedule. Constant a bit better than cosine, and has advantage for analysis'})
    warmup_ratio: float = field(default=0.03, metadata={"help": 'Fraction of steps to do a warmup for'})
    logging_steps: int = field(default=10, metadata={"help": 'The frequency of update steps after which to log the loss'})
    group_by_length: bool = field(default=False, metadata={"help": 'Group sequences into batches with same length. Saves memory and speeds up training considerably.'})
    save_strategy: str = field(default='epoch', metadata={"help": 'When to save checkpoints'})
    save_steps: int = field(default=250, metadata={"help": 'How often to save a model'})