trinity.common.rewards package

Submodules

• trinity.common.rewards.accuracy_reward module
  • AccuracyReward
    • AccuracyReward.__init__()
• trinity.common.rewards.agents_reward module
• trinity.common.rewards.base module
• trinity.common.rewards.composite_reward module
• trinity.common.rewards.format_reward module
  • FormatReward
    • FormatReward.__init__()
• trinity.common.rewards.human_reward module
• trinity.common.rewards.reward_fn module
  • RewardFn
    • RewardFn.__init__()
  • RMGalleryFn
    • RMGalleryFn.__init__()
• trinity.common.rewards.tool_reward module

Module contents

Reward functions for RFT

class trinity.common.rewards.RewardFn(**kwargs)

Bases: ABC

Base Reward Function Class.

abstract __init__(**kwargs) → None

class trinity.common.rewards.RMGalleryFn(reward_name, **kwargs)

Bases: RewardFn

Reward Function from RMGallery. https://github.com/modelscope/RM-Gallery

__init__(reward_name, **kwargs)

class trinity.common.rewards.AccuracyReward(answer_parser: Callable[[str], str] | None = None)

Bases: RewardFn

A reward function that rewards correct answers. Ref: https://github.com/huggingface/open-r1/blob/main/src/open_r1/rewards.py

__init__(answer_parser: Callable[[str], str] | None = None)

class trinity.common.rewards.CountDownRewardFn

Bases: RewardFn

A reward function that rewards for countdown task. Ref: Jiayi-Pan/TinyZero verl/utils/reward_score/countdown.py

__init__()

class trinity.common.rewards.FormatReward(pattern: str | None = None)

Bases: RewardFn

A reward function that checks if the reasoning process is enclosed within <think> and </think> tags, while the final answer is enclosed within <answer> and </answer> tags. Ref: https://github.com/huggingface/open-r1/blob/main/src/open_r1/rewards.py

__init__(pattern: str | None = None)

class trinity.common.rewards.MathRewardFn(answer_parser=<function simple_answer_parser>, pattern='.*?<think>.*?</think>\\s*<answer>.*?</answer>\\s*$')

Bases: RewardFn

A reward function that rewards for math task.

DEFAULT_FORMAT_PATTERN = '.*?<think>.*?</think>\\s*<answer>.*?</answer>\\s*$'

DEFAULT_ANSWER_PARSER() → str

__init__(answer_parser=<function simple_answer_parser>, pattern='.*?<think>.*?</think>\\s*<answer>.*?</answer>\\s*$') → None

class trinity.common.rewards.MathBoxedRewardFn(**kwargs)

Bases: RewardFn

A reward function that rewards for math task.

__init__(**kwargs) → None

class trinity.common.rewards.MathDAPORewardFn(enable_overlong_penalty: bool | None = None, penalty_factor: float | None = None, max_response_length: int | None = None, cache_length: int | None = None)

Bases: RewardFn

A reward function that follows the definition in DAPO for math task.

__init__(enable_overlong_penalty: bool | None = None, penalty_factor: float | None = None, max_response_length: int | None = None, cache_length: int | None = None) → None

compute_overlong_penalty(response_token)