GipFormer + Qwen32B integrate ACH audio-context checkpoint export.

This export contains trainable adapter components only:

Base components are loaded from the training config: frozen GipFormer encoder and Qwen/Qwen3-32B text model. ACH context is inserted before audio tokens with 8 context tokens.

Metrics: PW733-BTN WER 18.50%, Golden ASR WER 40.32%, non-command voice/noise acc 69.78%

Local source checkpoint: exp/gipformer_qwen32b_integrate_step2000_pw733_400/checkpoints/checkpoint_epoch1_step400

Evaluation summary: eval_summary.json