from dataclasses import dataclass, field from typing import Optional @dataclass class RolloutArguments: """ Arguments for rollout (on-policy distillation).""" def __post_init__(self): if self.eval_math_top_p is not None and not 0 < self.eval_math_top_p <= 1: raise ValueError('eval_math_top_p must be in (0,1] or None') rollout_num_engines: int = field( default=0, metadata={"help": "The number of engines for rollout."} ) rollout_tp_size: int = field( default=1, metadata={"help": "Tensor parallel size for each rollout engine."} ) rollout_engine_concurrency: int = field( default=512, metadata={"help": "Maximum concurrent requests per rollout engine."} ) rollout_enable_sleep: bool = field( default=False, metadata={"help": "Enable sleep mode for rollout engines."} ) rollout_mem_fraction_static: float = field( default=0.6, metadata={"help": "GPU memory utilization for each rollout engine."} ) top_p: float = field( default=1.0, metadata={"help": "Top-p sampling for rollout."} ) eval_math_top_p: float = field(default=None, metadata={'help':'MATH sampling top-p; None inherits rollout top_p.'}) temperature: float = field( default=1.0, metadata={"help": "Temperature for rollout."} ) n_samples_per_prompt: int = field( default=1, metadata={"help": "Sample n responses per prompt."} ) rollout_batch_size: int = field( default=32, metadata={"help": "Number of prompts for each rollout."} ) generate_max_len: int = field( default=2048, metadata={"help": "Max generation tokens during rollout."} ) print_rollout_sample: bool = field( default=False, metadata={"help": "Whether to print a rollout sample after each rollout."} ) alfworld_config: Optional[str] = field( default=None, metadata={ "help": "Path to an ALFWorld YAML config. When set, on-policy KD " "collects multi-turn student trajectories from AlfredTWEnv." }, ) alfworld_protocol: str = field( default="kdflow", metadata={ "help": "ALFWorld training rollout protocol: 'kdflow' keeps the " "original Thought/Action chat history, while 'verl_agent' " "uses the migrated single-prompt / protocol." }, ) alfworld_split: str = field( default="train", metadata={ "help": "ALFWorld split: train, eval_in_distribution, or " "eval_out_of_distribution." }, ) alfworld_max_turns: int = field( default=30, metadata={"help": "Maximum assistant/environment turns per ALFWorld episode."}, ) alfworld_episodes_per_epoch: int = field( default=1024, metadata={"help": "Number of ALFWorld episodes sampled per training epoch."}, ) alfworld_eval_episodes: int = field( default=128, metadata={ "help": "Number of episodes for standalone ALFWorld evaluation." }, ) alfworld_eval_protocol: str = field( default="kdflow", metadata={ "help": "Standalone ALFWorld evaluation protocol. Use 'kdflow' " "for the original KDFlow prompt or 'verl_agent' for the " "protocol migrated from langfengq/verl-agent." }, ) alfworld_eval_seeds: str = field( default="123,456,789", metadata={ "help": "Comma-separated script-level seeds used by the " "verl-agent-compatible evaluator." }, ) alfworld_eval_seed_offset: int = field( default=1000, metadata={ "help": "Offset added to each script-level seed for the validation " "environment. verl-agent uses 1000." }, ) alfworld_eval_resume: bool = field( default=True, metadata={ "help": "Reuse completed, configuration-compatible per-seed " "ALFWorld evaluation results. Set False to force rerunning " "every seed." }, ) alfworld_history_length: int = field( default=4, metadata={ "help": "Number of recent observation/action pairs included by the " "verl-agent ALFWorld prompt." }, ) alfworld_admissible_limit: int = field( default=30, metadata={ "help": "Maximum number of ALFWorld admissible actions shown per turn; " "use 0 to show all actions." }, )