LabVLA-5B-Base appears to miss norm_stats.json required for correct deployment

[eggplantsalt]

Hi, thanks for releasing LabVLA.

I downloaded zjunlp/LabVLA-5B-Base and deployed it with the official
deployment/serve_labvla.py.
The public HF checkpoint contains model.safetensors, config.json,
vlm_config.json and tokenizer/processor files, but it does not contain:

• norm_stats.json
• labvla_schema.json / labvla_schemas.json
• training_state/training_state.pt

The current checkpoint saver appears to save the model/VLM assets under
checkpoint/pretrained_model/, while norm_stats.json and schema files are
saved at the checkpoint root.
Could you please release the exact norm_stats.json used by
LabVLA-5B-Base, or preferably the full checkpoint root used for deployment?

Thanks!

[BaochangRen]

Hi, thank you very much for your interest in LabVLA and for the detailed report!

You are right: norm_stats.json, labvla_schema.json and training_state/training_state.pt are not in zjunlp/LabVLA-5B-Base. They are written by scripts/train.py at the root of each outputs/<job>/checkpoint-N/ during (fine-)training, and they are dataset-specific (normalization statistics and the state/action layout of the training data), so the base model itself does not carry them. LabVLA-5B-Base is the post-trained checkpoint meant to be fine-tuned on your task data; the resulting checkpoint-N/ contains all of these files and can be deployed directly with PRETRAINED_PATH=/path/to/checkpoint-N bash deployment/deploy.sh.

Our README Quick Start is misleading on this point (it suggests deploying the base directly); we will fix it. Sorry for the confusion, and please feel free to reach out if you have any further questions!

[eggplantsalt]

Hi, thank you very much for your interest in LabVLA and for the detailed report!

You are right: norm_stats.json, labvla_schema.json and training_state/training_state.pt are not in zjunlp/LabVLA-5B-Base. They are written by scripts/train.py at the root of each outputs/<job>/checkpoint-N/ during (fine-)training, and they are dataset-specific (normalization statistics and the state/action layout of the training data), so the base model itself does not carry them. LabVLA-5B-Base is the post-trained checkpoint meant to be fine-tuned on your task data; the resulting checkpoint-N/ contains all of these files and can be deployed directly with PRETRAINED_PATH=/path/to/checkpoint-N bash deployment/deploy.sh.

Our README Quick Start is misleading on this point (it suggests deploying the base directly); we will fix it. Sorry for the confusion, and please feel free to reach out if you have any further questions!

OK，thx for reply