Qwen3-ASR / MoulSot Fixed Full Training Pack
This version fixes the torchcodec dataset problem by using:
Then it reads audio bytes manually with soundfile, so dataset preparation does not require TorchCodec.
One-command flow
Run:
bash RUN_ALL_STEP_BY_STEP.shIt will do:
- Install stable dependencies.
- Login to Hugging Face.
- Prepare a 20-row test dataset.
- Verify the JSONL is not empty.
- Prepare the full MoulSot dataset.
- Download the official Qwen3-ASR fine-tuning script.
- Start training with autosave/resume.
Resume training
If training stops, run:
SAVE_STEPS=25 bash 04_train_autosave_resume.shIt automatically resumes from the latest checkpoint.
Important
No script can save after a hard GPU kill or power loss. The protection is frequent checkpointing:
