Skip to content

vllm.reasoning.seedoss_reasoning_parser

SeedOSSReasoningParser

Bases: BaseThinkingReasoningParser

Reasoning parser for SeedOSS model.

The SeedOSS model uses ... tokens to denote reasoning content text. This parser extracts the reasoning content from the model output. Similar to DeepSeek R1, it supports cases where the model doesn't generate the start token.

Source code in vllm/reasoning/seedoss_reasoning_parser.py
@ReasoningParserManager.register_module("seed_oss")
class SeedOSSReasoningParser(BaseThinkingReasoningParser):
    """
    Reasoning parser for SeedOSS model.

    The SeedOSS model uses <seed:think>...</seed:think> tokens to 
    denote reasoning content text. This parser extracts 
    the reasoning content from the model output.
    Similar to DeepSeek R1, it supports cases 
    where the model doesn't generate the start token.
    """

    @property
    def start_token(self) -> str:
        """The token that starts reasoning content."""
        return "<seed:think>"

    @property
    def end_token(self) -> str:
        """The token that ends reasoning content."""
        return "</seed:think>"

end_token property

end_token: str

The token that ends reasoning content.

start_token property

start_token: str

The token that starts reasoning content.