---
license: apache-2.0
library_name: laya
base_model: convaiinnovations/laya
tags: [laya, system-one, android, phone-agent, decision-model]
---

# laya-android-control

Finetune of [convaiinnovations/laya](https://huggingface.co/convaiinnovations/laya) for **Android phone control**.

## Training data

- Source: [eros438/phone-decisions-ac](https://www.kaggle.com/datasets/eros438/phone-decisions-ac).
- That file holds 18,354 decision rows.
- The rows convert [xwm/AndroidControl](https://huggingface.co/datasets/xwm/AndroidControl).
- Training ran 4 epochs on a Kaggle T4 with `laya-train` defaults.

## Results

| Model | test operations accuracy |
|---|---|
| Base Laya-english | 0.261 |
| Head-only finetune | 0.574 |
| Full finetune | 0.665 |

In-sample calibration gave accuracy 0.85 and ECE 0.037.

## Use

```python
import laya
agent = laya.load("arnabwithab/laya-android-control", device="cuda")
agent.predict("Goal: ...", {"operation": {"type": "choice",
    "instructions": "Which operation moves this phone task forward?",
    "criteria": {"TAP": "...", "TYPE": "...", "SCROLL": "...",
                 "BACK": "...", "WAIT": "...", "DONE": "..."}}})
```

Serve it with `laya-serve` pointed at this repo.