Skip to content

Question about state token in Pi0.5 VLASH #21

@jibby2803

Description

@jibby2803

Hi, I am applying SharedObservation training on Pi0.5. However, I see that the state are processing differently compared to original pi0.5.

In original Pi0.5, state is discretized and treated as text tokens and passed to VLM as a part of prefix_embs. However, when training with Shared observation config, I see that state is not processed in VLM and it is processed like a suffix embed.

Image Image

Can anyone help me to clarify it? Is my understand correct?
If yes, How does this difference effect the model performance?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions