About training Plus-Vanilla

[jee019]

I'm trying to reproduce training for the repository and have three questions plus a reproducible problem I'm hitting when running the training script.

1. finetune_sensevoice_glm4voice — is this the script used to train VITA-Audio-Plus-Vanilla (the baseline version)? If so, does VITA-Audio-Plus-Vanilla intentionally not use MCTP, meaning stage2 and stage3 are skipped for that configuration?
2. When I run the training script locally I see the model input is not being passed correctly and training fails. Is this a known bug or misconfiguration? Could you confirm whether the repo's training scripts are expected to work as-is, or if there are required pre-processing / arguments not documented in the README?
3. What is the precise role of the --dataset_joint option? I couldn't find a short explanation in the README or scripts.

Thanks in advance — happy to provide any further details or small reproducer files if helpful.