Zerith-Robotics commited on
Commit
36bf4ff
·
verified ·
1 Parent(s): ad16596

Align model card naming with ZERITH 2.5

Browse files
Files changed (1) hide show
  1. README.md +3 -3
README.md CHANGED
@@ -3,12 +3,12 @@ tags:
3
  - robotics
4
  - safetensors
5
  ---
6
- # Z-VLA-LIBERO
7
 
8
- Z-VLA-LIBERO policy weights. The deployment manifest contains the architecture, action/camera interfaces and required normalization. Use the settings in inference.json.
9
 
10
  The learned tensors retain their original FP32 values. Weights are stored in `model.safetensors`; file size and SHA-256 are recorded in `manifest.json`.
11
 
12
  External VAE, geometry and text-encoder assets must be obtained from the upstream revisions in `dependencies.json`. Text weights and precomputed text embeddings are not included. Generate instruction embeddings using that pinned encoder and the declared token/padding contract.
13
 
14
- These are Z-VLA artifacts, not Transformers AutoModel checkpoints. Use the matching Z-VLA neural modules and action interface. A full benchmark rerun with newly generated text embeddings is required before claiming numerical or benchmark equivalence to any previous evaluation.
 
3
  - robotics
4
  - safetensors
5
  ---
6
+ # ZERITH 2.5 LIBERO
7
 
8
+ ZERITH 2.5 LIBERO policy weights. The deployment manifest contains the architecture, action/camera interfaces and required normalization. Use the settings in inference.json.
9
 
10
  The learned tensors retain their original FP32 values. Weights are stored in `model.safetensors`; file size and SHA-256 are recorded in `manifest.json`.
11
 
12
  External VAE, geometry and text-encoder assets must be obtained from the upstream revisions in `dependencies.json`. Text weights and precomputed text embeddings are not included. Generate instruction embeddings using that pinned encoder and the declared token/padding contract.
13
 
14
+ These are ZERITH 2.5 artifacts, not Transformers AutoModel checkpoints. Use the matching ZERITH 2.5 neural modules and action interface. A full benchmark rerun with newly generated text embeddings is required before claiming numerical or benchmark equivalence to any previous evaluation.