Class GRPOScriptArguments

rft/configs.py:22–42

Script arguments for the GRPO training script. Args: reward_funcs (`list[str]`): List of reward functions. Possible values: 'accuracy', 'format'.

@dataclass
class GRPOScriptArguments(ScriptArguments):
    """
    Script arguments for the GRPO training script.

    Args:
        reward_funcs (`list[str]`):
            List of reward functions. Possible values: 'accuracy', 'format'.
    """

    reward_funcs: list[str] = field(
        default_factory=lambda: ["accuracy", "format"],
        metadata={"help": "List of reward functions. Possible values: 'accuracy', 'format'"},
    )
    max_line_res: Optional[int] = field(
        default=None,
        metadata={"help": "Maximum number of pixels for the longest line of the image"},
    )
    eval_dataset_name: Optional[str] = field(
        default=None,
        metadata={"help": "Eval dataset name or path"}
    )


@dataclass