Neo AuT–Qwen projector (LibriSpeech ~10h)

Training

• Encoder: Qwen/Qwen3-Omni-30B-A3B-Instruct AuT tower only (thinker.audio_tower.*)
• Data: LibriSpeech clean train.100 (2780 utts ≈ 10h), 4 epochs, 8×H200
• Final eval (100 samples): normalized WER 33.69%, eval_loss 0.0702
• WandB: aut-qwen-librispeech-10h-h200x8

Files

• projector.pt / projector_config.json / training_config.yaml
• audio_start_embedding.pt
• tokenizer/