nemo_rl.data.chat_utils#
Message selection validation for chat data processors.
Module Contents#
Functions#
Validate turn selection, defaulting to all assistant messages. |
API#
- nemo_rl.data.chat_utils.normalize_message_loss_mask(
- messages: Sequence[dict[str, Any]],
- message_loss_mask: Sequence[int] | None,
Validate turn selection, defaulting to all assistant messages.
Supplied masks must have one binary integer per message and may select only assistant turns. This utility preserves earlier assistant selections, but the chat collator supervises only the final assistant message: earlier entries never enable supervision. A zero never removes a message from the rendered conversation.