FunASRNanoTrainingStage

Enums · AiDotNet.Enums

Which modules a Fun-ASR-nano training step updates: the report’s supervised fine-tuning stages.

How It Works

The Fun-ASR report (arXiv:2509.12508, §4.2) fine-tunes in stages. Stage 1 trains only the adaptor. Stage 2 also trains the audio encoder. Stage 3 freezes both and trains LoRA adapters on the LLM. Stage 4 trains the encoder, the adaptor and the LoRA adapters together. Stage 5 adds a CTC decoder on the frozen encoder and trains only that. The LLM’s own weights are never updated: every stage presupposes a pretrained Qwen3. Joint is the default because a model built here starts from random weights, which those stages would leave frozen; it is NOT the report’s recipe.

Fields

Field Summary
FunASRNanoTrainingStage AdaptorAlignment Stage 1: only the adaptor, aligning the encoder’s output with the LLM’s embedding space.
FunASRNanoTrainingStage CtcDecoder Stage 5: only the CTC decoder on the frozen encoder; targets are ids in the CTC vocabulary.
FunASRNanoTrainingStage EncoderAndAdaptor Stage 2: the audio encoder and the adaptor.
FunASRNanoTrainingStage FullFineTuning Stage 4: the encoder, the adaptor and the LLM’s LoRA adapters together.
FunASRNanoTrainingStage Joint Every parameter, the LLM’s and the CTC decoder’s included, in one step.
FunASRNanoTrainingStage LlmLora Stage 3: only the LLM’s LoRA adapters; the encoder and the adaptor are frozen.
int value__