Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -7,17 +7,30 @@ tags: [laya, system-one, android, phone-agent, decision-model]
|
|
| 7 |
|
| 8 |
# laya-android-control
|
| 9 |
|
| 10 |
-
|
| 11 |
-
(ModernBERT-large, 421M) for Android phone control.
|
| 12 |
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
(a parquet mirror of Google's AndroidControl: 15,283 human demos, 833 apps).
|
| 16 |
-
Each step became an `operation` choice (TAP/TYPE/SCROLL/BACK/WAIT/DONE, state = goal + step
|
| 17 |
-
history) plus a `done` noul. 4 epochs on a Kaggle T4 via `laya-train` (defaults).
|
| 18 |
|
| 19 |
-
|
| 20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
|
| 22 |
```python
|
| 23 |
import laya
|
|
@@ -28,4 +41,4 @@ agent.predict("Goal: ...", {"operation": {"type": "choice",
|
|
| 28 |
"BACK": "...", "WAIT": "...", "DONE": "..."}}})
|
| 29 |
```
|
| 30 |
|
| 31 |
-
|
|
|
|
| 7 |
|
| 8 |
# laya-android-control
|
| 9 |
|
| 10 |
+
## What this is
|
|
|
|
| 11 |
|
| 12 |
+
- Finetune of [convaiinnovations/laya](https://huggingface.co/convaiinnovations/laya) for Android phone control.
|
| 13 |
+
- Base model: ModernBERT-large, 421M parameters.
|
|
|
|
|
|
|
|
|
|
| 14 |
|
| 15 |
+
## Training data
|
| 16 |
+
|
| 17 |
+
- Source: [eros438/phone-decisions-ac](https://www.kaggle.com/datasets/eros438/phone-decisions-ac).
|
| 18 |
+
- That file holds 18,354 decision rows.
|
| 19 |
+
- The rows convert [xwm/AndroidControl](https://huggingface.co/datasets/xwm/AndroidControl).
|
| 20 |
+
- That dataset mirrors Google AndroidControl.
|
| 21 |
+
- Google AndroidControl holds 15,283 human demonstrations in 833 apps.
|
| 22 |
+
- Each step became one operation choice: TAP, TYPE, SCROLL, BACK, WAIT, or DONE.
|
| 23 |
+
- State = goal + step history.
|
| 24 |
+
- Each step also got one done check.
|
| 25 |
+
- Training ran 4 epochs on a Kaggle T4 with `laya-train` defaults.
|
| 26 |
+
|
| 27 |
+
## Results
|
| 28 |
+
|
| 29 |
+
- Held-out goals: 984.
|
| 30 |
+
- Operation accuracy rose from 0.261 to 0.665.
|
| 31 |
+
- In-sample calibration gave accuracy 0.85 and ECE 0.037.
|
| 32 |
+
|
| 33 |
+
## Use
|
| 34 |
|
| 35 |
```python
|
| 36 |
import laya
|
|
|
|
| 41 |
"BACK": "...", "WAIT": "...", "DONE": "..."}}})
|
| 42 |
```
|
| 43 |
|
| 44 |
+
Serve it with `laya-serve` pointed at this repo.
|