add: initialize domain-adapted model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: transformers
|
| 3 |
+
pipeline_tag: robotics
|
| 4 |
+
license: other
|
| 5 |
+
base_model: argo11/parc2026-openvla-oft-plus-base
|
| 6 |
+
datasets:
|
| 7 |
+
- argo11/openvla-oft-plus-robustness-recovery
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# OpenVLA-OFT+ Domain-Adapted Checkpoints
|
| 11 |
+
|
| 12 |
+
This repository stores independently trained OpenVLA-OFT+ adapters and their
|
| 13 |
+
action-head and proprio-projector components.
|
| 14 |
+
|
| 15 |
+
No production or submission candidate is published on `main` yet. Bounded
|
| 16 |
+
environment-validation checkpoints are kept on branches whose names start with
|
| 17 |
+
`smoke-colab-20step`.
|
| 18 |
+
|
| 19 |
+
## Provenance
|
| 20 |
+
|
| 21 |
+
- Base model: `argo11/parc2026-openvla-oft-plus-base`
|
| 22 |
+
- Base revision: `0d923bf7af68b7eac9290cdd65f597cad7b48596`
|
| 23 |
+
- Training dataset: `argo11/openvla-oft-plus-robustness-recovery`
|
| 24 |
+
- Dataset revision: `c3807717b10ca28c56630508262af7163ca59d45`
|
| 25 |
+
- OpenVLA-OFT code revision: `e4287e94541f459edc4feabc4e181f537cd569a8`
|
| 26 |
+
|
| 27 |
+
The dataset is published with `license: other`; its upstream-specific terms are
|
| 28 |
+
not fully resolved. Public availability does not by itself grant unrestricted
|
| 29 |
+
reuse. See the dataset card and third-party notices before using these weights.
|
| 30 |
+
The pinned base Model Card declares the base weights under MIT. Because the
|
| 31 |
+
training dataset's aggregate license remains unresolved, this repository also
|
| 32 |
+
uses `license: other`; the base-model MIT declaration does not resolve data or
|
| 33 |
+
third-party asset rights for the derived checkpoint.
|
| 34 |
+
|
| 35 |
+
No related online W&B run is claimed for the Colab smoke validation. Training
|
| 36 |
+
uses W&B offline mode unless a later release documents another provenance path.
|