verl/trainer/main_ppo.py

• _select_rm_score_fn · function · L67-L79 — def _select_rm_score_fn(data_source, reward_type: str = "em")
• RewardManager · class · L82-L547 — class RewardManager()
• init · method · L86-L135 — def init( self, tokenizer, num_examine, format_score: float = FORMAT_SCORE, reward_type: str = REWARD_TYPE, reward_mode: str = REWARD_MODE, no_search_penalty: float = NO_SEARCH_PENALTY, turn_free: int = TURN_FREE, turn_penalty_per_turn: float = TURN_PENALTY_PER_TURN, no_answer_penalty: float = NO_ANSWER_PENALTY, curriculum_step: int = CURRICULUM_STEP, # confidence 统计相关参数 confidence_mode: str = CONFIDENCE_MODE, correct_threshold: float = CORRECT_THRESHOLD, # 评估模式：final_score 只用 EM（base_score），不叠加任何惩罚 eval_only_em: bool = False, ) -> None
• call · method · L137-L511 — def call(self, data: DataProto)
• _extract_answer_with_format_check · method · L514-L522 — def _extract_answer_with_format_check(self, text: str) -> tuple
• _is_no_answer · method · L524-L545 — def _is_no_answer(self, ans: str) -> bool
• main · function · L555-L560 — def main(config)
• main_task · function · L564-L672 — def main_task(config)