qwen35-v46-audio-graft-koen-vaqa

Architecture

End-to-end status

• Inference path is end-to-end.
• Default training is projector-only for Kaggle stability.
• Qwen and Whisper are frozen by default.

Components

• Base LLM: Qwen/Qwen3.5-9B
• Text LoRA: Hellohihihih/qwen35-v30-textmix-lora
• Audio encoder: openai/whisper-small
• Trainable module: audio_projector