ConversationSourceConfig

Module: fast_llm.data.preparation.gpt_memmap.config

Variant of: LanguageModelSourceConfig — select with type: conversation

Inherits from: LanguageModelSourceConfig

Fields

messages — core

Type: str    Default: "messages"

Field containing the conversation messages list. Each message should have 'role' and 'content' keys.

train_on_eos — optional

Type: bool    Default: False

Include the end-of-sequence token appended after the final message in the training loss. When disabled, that token is masked from the loss.