SpeechLlmTrainingScope

Enums · AiDotNet.Enums

Which parameters a training step of an LLM-based speech recognizer (a speech encoder, an adapter and a pretrained LLM) updates.

How It Works

The papers behind these models fine-tune from a pretrained LLM: they keep its weights fixed and train the encoder, the adapter and LoRA adapters on the LLM (EncoderAdapterAndLora). A model built here starts from random weights; a frozen random LLM, whose output head cannot sharpen its predictions, would leave nothing able to fit a transcript, so AllParameters is the default. It is NOT the papers’ recipe.

Fields

Field Summary
SpeechLlmTrainingScope AllParameters Every parameter, the LLM’s included.
SpeechLlmTrainingScope EncoderAdapterAndLora The papers’ recipe: the encoder, the adapter and the LLM’s LoRA adapters; the LLM’s own weights stay fixed.
int value__