Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- provael
|
| 5 |
+
- vla
|
| 6 |
+
- red-teaming
|
| 7 |
+
- leaderboard
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# Provael leaderboard — submission queue
|
| 11 |
+
|
| 12 |
+
Incoming results for the [Provael VLA policy ASR leaderboard](https://www.provael.com/leaderboard/).
|
| 13 |
+
This dataset is a **queue**, not the board. A maintainer validates each submission and promotes it;
|
| 14 |
+
until then a file here is a request, not a published result.
|
| 15 |
+
|
| 16 |
+
## How a result gets here
|
| 17 |
+
|
| 18 |
+
Either route works and both end in the same review:
|
| 19 |
+
|
| 20 |
+
1. The **Submit a result** tab on the
|
| 21 |
+
[Space](https://huggingface.co/spaces/Sattyam/provael-leaderboard) — uploads your JSON and opens
|
| 22 |
+
a pull request against this dataset.
|
| 23 |
+
2. A [GitHub issue](https://github.com/provael/provael/issues/new) with the JSON attached. Needs no
|
| 24 |
+
Hugging Face account, and is the route the Space falls back to when this queue is unavailable.
|
| 25 |
+
|
| 26 |
+
## What a submission must contain
|
| 27 |
+
|
| 28 |
+
The output of `provael leaderboard build` — which carries the ASR, its 95% Wilson interval, the
|
| 29 |
+
benign control, the attempt count, and the transfer status of every row. Validate before sending:
|
| 30 |
+
|
| 31 |
+
```bash
|
| 32 |
+
pip install provael
|
| 33 |
+
provael submit --dry-run --in <your-run-dir>
|
| 34 |
+
```
|
| 35 |
+
|
| 36 |
+
That signs and prints exactly what would be submitted, touching no network.
|
| 37 |
+
|
| 38 |
+
## What gets rejected, so it is not a surprise
|
| 39 |
+
|
| 40 |
+
- A result with no benign control. The control is what makes an ASR mean anything.
|
| 41 |
+
- A row that reports an attack success rate without its denominator.
|
| 42 |
+
- A stub-policy run presented as a real-model result. Stub runs are welcome and are labelled as
|
| 43 |
+
stub — the labelling is the requirement, not the policy.
|
| 44 |
+
|
| 45 |
+
Null results are **wanted**. The published board already carries measured zeros, and a family that
|
| 46 |
+
does not transfer is a finding.
|