trinity.algorithm.algorithm module

Algorithm classes.

class trinity.algorithm.algorithm.ConstantMeta(name, bases, namespace, /, **kwargs)

基类：ABCMeta

class trinity.algorithm.algorithm.AlgorithmType

基类：ABC

use_critic: bool

use_reference: bool

compute_advantage_in_trainer: bool

can_balance_batch: bool

schema: str

abstractmethod classmethod default_config() → Dict

classmethod name() → str

classmethod check_config(config: Config) → None

class trinity.algorithm.algorithm.SFTAlgorithm

基类：AlgorithmType

SFT Algorithm.

use_critic: bool = False

use_reference: bool = False

compute_advantage_in_trainer: bool = False

can_balance_batch: bool = True

schema: str = 'sft'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.PPOAlgorithm

基类：AlgorithmType

PPO Algorithm.

use_critic: bool = True

use_reference: bool = True

compute_advantage_in_trainer: bool = True

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.GRPOAlgorithm

基类：AlgorithmType

GRPO algorithm.

use_critic: bool = False

use_reference: bool = True

compute_advantage_in_trainer: bool = False

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.ReinforcePlusPlusAlgorithm

基类：AlgorithmType

Reinforce++ algorithm.

use_critic: bool = False

use_reference: bool = True

compute_advantage_in_trainer: bool = True

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.RLOOAlgorithm

基类：AlgorithmType

RLOO algorithm.

use_critic: bool = False

use_reference: bool = True

compute_advantage_in_trainer: bool = True

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.OPMDAlgorithm

基类：AlgorithmType

OPMD algorithm.

use_critic: bool = False

use_reference: bool = True

compute_advantage_in_trainer: bool = False

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.AsymREAlgorithm

基类：AlgorithmType

AsymRE algorithm.

use_critic: bool = False

use_reference: bool = False

compute_advantage_in_trainer: bool = False

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.DPOAlgorithm

基类：AlgorithmType

DPO algorithm.

use_critic: bool = False

use_reference: bool = True

compute_advantage_in_trainer: bool = False

can_balance_batch: bool = False

schema: str = 'dpo'

classmethod default_config() → Dict

classmethod check_config(config: Config) → None

class trinity.algorithm.algorithm.TOPRAlgorithm

基类：AlgorithmType

TOPR algorithm. See https://arxiv.org/pdf/2503.14286v1

use_critic: bool = False

use_reference: bool = True

compute_advantage_in_trainer: bool = False

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.CISPOAlgorithm

基类：AlgorithmType

CISPO algorithm. See https://arxiv.org/abs/2506.13585

use_critic: bool = False

use_reference: bool = True

compute_advantage_in_trainer: bool = False

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.GSPOAlgorithm

基类：AlgorithmType

GSPO algorithm. See https://arxiv.org/pdf/2507.18071

use_critic: bool = False

use_reference: bool = True

compute_advantage_in_trainer: bool = False

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.SAPOAlgorithm

基类：AlgorithmType

SAPO (Soft Adaptive Policy Optimization) algorithm.

SAPO uses a smooth, temperature-controlled soft gate instead of hard clipping to stabilize training while maintaining effective learning.

use_critic: bool = False

use_reference: bool = True

compute_advantage_in_trainer: bool = False

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.MIXAlgorithm

基类：AlgorithmType

MIX algorithm.

use_critic: bool = False

use_reference: bool = True

compute_advantage_in_trainer: bool = False

use_rollout: bool = True

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.MIXCHORDAlgorithm

基类：AlgorithmType

MIX algorithm.

use_critic: bool = False

use_reference: bool = True

compute_advantage_in_trainer: bool = False

use_rollout: bool = True

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.RAFTAlgorithm

基类：AlgorithmType

RAFT Algorithm. This algorithm is conceptually similar to Supervised Fine-Tuning (SFT) but is designed to work with experience schema from rollouts.

use_critic: bool = False

use_reference: bool = False

compute_advantage_in_trainer: bool = False

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.sPPOAlgorithm

基类：AlgorithmType

sPPO Algorithm.

use_critic: bool = False

use_reference: bool = False

compute_advantage_in_trainer: bool = False

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.RECAlgorithm

基类：AlgorithmType

REC Algorithm.

use_critic: bool = False

use_reference: bool = True

compute_advantage_in_trainer: bool = False

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.MultiStepGRPOAlgorithm

基类：AlgorithmType

Multi-Step GRPO Algorithm.

use_critic: bool = False

use_reference: bool = True

compute_advantage_in_trainer: bool = False

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict

class trinity.algorithm.algorithm.OnPolicyDistillAlgorithm

基类：AlgorithmType

On-Policy Distillation Algorithm.

Reference: Tinker library.

Workflow stores teacher_logprobs in experience.info["teacher_logprobs"]. Trainer's advantage_fn computes: advantages = teacher_logprobs - student_logprobs Trainer uses:

importance_sampling loss if no clipping is needed ppo loss if clipping is needed, for better stability

use_critic: bool = False

use_reference: bool = False

compute_advantage_in_trainer: bool = True

can_balance_batch: bool = True

schema: str = 'experience'

classmethod default_config() → Dict