verl/trainer/conf_adv_config.py

• get_lsc_alpha · function · L32-L39 — def get_lsc_alpha(cur_step: int) -> float
• get_fc_alpha · function · L42-L49 — def get_fc_alpha(cur_step: int) -> float
• compute_calib_reward_stake · function · L56-L81 — def compute_calib_reward_stake(q: np.ndarray, c: np.ndarray) -> np.ndarray
• compute_fc_consistency_reward · function · L84-L112 — def compute_fc_consistency_reward( c_fc: np.ndarray, c_lsc: np.ndarray, threshold: float = FC_CONSISTENCY_THRESHOLD, ) -> np.ndarray
• compute_fc_rank_consistency_reward · function · L115-L155 — def compute_fc_rank_consistency_reward( c_fc: np.ndarray, c_lsc: np.ndarray, delta: float = FC_RANK_DELTA, ) -> np.ndarray
• _sgn · function · L135-L141 — def _sgn(a, b, d)
• apply_dual_calib_to_scores · function · L162-L228 — def apply_dual_calib_to_scores( scores: "torch.Tensor", # (batch_size,) 原始 outcome score（q） lsc_confs: "torch.Tensor", # (batch_size,) last_step_confidences，-1 表示无效 fc_confs: "torch.Tensor", # (batch_size,) final_confidences，-1 表示无效 uid_list: list, cur_step: int, ) -> "torch.Tensor"
• compute_ece · function · L235-L267 — def compute_ece(confidences: list, is_correct: list, n_bins: int = 20)
• compute_auroc · function · L270-L290 — def compute_auroc(confidences, is_correct)