File size: 2,851 Bytes
a917021
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
---
pretty_name: WeightSpace research artifacts
tags:
- robotics
- openvla
- libero
- model-editing
---

# WeightSpace research artifacts

This release contains **317 checkpoint/figure artifacts (335.5 MiB)**
plus sanitized result records in `metadata/data/`. The upload is limited to
checkpoints and figures. Trajectories, videos, raw sessions, account files,
environments and intermediate inference caches are not included.

## OpenVLA base checkpoint

The unmodified 14 GiB base model already lives at
[openvla/openvla-7b-finetuned-libero-spatial](https://huggingface.co/openvla/openvla-7b-finetuned-libero-spatial/tree/962318cec55ac10993ff0f5f43eda9a270b4c873).
It is pinned to revision `962318cec55ac10993ff0f5f43eda9a270b4c873` and is not duplicated here.
The four frozen sparse patches are under
`artifacts/openvla_study/results/attempts/<attempt>/submitted.npz`.
Apply each independently to that same original checkpoint; the empty Opus
three-task patch leaves it unchanged. The exact base-file hashes are in
`metadata/data/openvla/checkpoint_manifest.json`.

## Measured OpenVLA results

| Policy | Overall successes | Selected-task successes |
|---|---|---|
| Baseline | 27/30 | Single 3/5; three-task 12/15 |
| Opus / single | 27/30 | 3/5 |
| Opus / three (no-op) | 27/30 | 12/15 |
| Fable / single | 24/30 | 3/5 |
| Fable / three | 26/30 | 12/15 |

Five final episodes per task, one attempt per condition. Three conditions had
infrastructure continuations after memory failures. Fable single-task received
no valid development feedback because of a renderer failure. These results do
not establish a model ranking. The hand/drawer study has separate metrics and
is documented in `metadata/data/robot_policy/`.

## Layout and integrity

- `artifacts/policies/`: canonical `.npz` MLP checkpoints.
- `artifacts/results/`: selected small-policy submissions and initial weights.
- `artifacts/openvla_study/results/`: the four frozen OpenVLA sparse patches.
- `artifacts/reports/figures/`: PNG, PDF and editable SVG figures.
- `metadata/data/`: sanitized configurations, scores, audits and base-model manifest.
- `artifact_manifest.json`: relative restoration paths, sizes and SHA-256 hashes.

The accompanying GitHub project's `scripts/download_artifacts.py` restores files
to those relative paths and verifies hashes. Raw recordings and private runtime state are not published.

## Licensing and provenance

No blanket license has been assigned to the original WeightSpace code/artifacts.
Upstream components keep their own terms: OpenVLA and LIBERO use MIT licenses;
the copied IsaacLab source files identify BSD-3-Clause. Preserved notices are in
`metadata/THIRD_PARTY_LICENSES/`. Converted NVIDIA checkpoints and derivatives
remain subject to their upstream terms. Do not treat this dataset card as a new
license grant for those materials.