AppGen Qwen3 policy SFT — arm R
This public checkpoint was trained only on user-owned synthetic AppGen data. It contains no AMEX, GUIOwl, AndroidWorld, ScreenSpot, RefExp, MoTIF, VWB, or other evaluation-benchmark examples.
- Initialization:
Qwen/Qwen3-VL-8B-Instructat0c351dd01ed87e9c1b53cbc748cba10e6187ff3b - Rows: 22,560
- Dataset SHA-256:
c0eb0f5c6fae9099df5301397534be7706d39e1d661852eb7050a8371c5a7b00 - Training type:
full - Learning rate:
3.5e-07 - Optimizer steps: 705
- Vision encoder and aligner: frozen
- Global batch size: 32
- Prompt: strict one-action
<think>+<tool_call>policy with ten-step semantic history
For evaluation, use the exact appgen_system_prompt.txt included in this repository,
enable explicit thinking, and provide normalized integer coordinates in [0, 999].