tinker_cookbook.renderers.TrainOnWhat
class tinker_cookbook.renderers.TrainOnWhat(StrEnum)
Enum controlling which parts of the sequence to compute loss on.
Members:
LAST_ASSISTANT_MESSAGE: Train only on the final assistant message.
LAST_ASSISTANT_TURN: Train on the last assistant turn (including any
tool calls and tool responses leading to the final assistant reply).
ALL_ASSISTANT_MESSAGES: Train on every assistant message in the conversation.
ALL_MESSAGES: Train on all messages regardless of role.
ALL_TOKENS: Train on every token in the sequence (including special tokens).
ALL_USER_AND_SYSTEM_MESSAGES: Train on user and system messages only.
CUSTOMIZED: Use per-message train flags from the dataset.