appgen-qwen3-thinktag-lr5e-7-ep1-continue-sft-norm1000-nav-only
Continue-SFT checkpoint of luca0621/appgen-qwen3-thinktag-lr5e-7-ep1 trained on
navigation-only GUI data in the mobile_use tool-call format.
- Base model:
luca0621/appgen-qwen3-thinktag-lr5e-7-ep1 - Output format:
mobile_usetool call with normalized 0-1000 coordinates - Dataset:
mobile_use_norm1000, navigation-only split - Training: 1 epoch, learning rate 5e-7, full fine-tuning, max length 4096
- Checkpoint:
checkpoint-2406
Real-World Eval
Evaluation was run with prompt style gelab, coordinate space 0-1000 point-in-bbox,
and fixed image pixels 2629536.
| Benchmark | Accuracy | Correct / Total |
|---|---|---|
| ScreenSpot | 90.33% | 1149 / 1272 |
| ScreenSpot-v2 | 93.40% | 1188 / 1272 |
| FuncPred | 62.00% | 1240 / 2000 |
| MoTIF | 76.36% | 966 / 1265 |
| Refexp | 86.90% | 491 / 565 |
| VWB-EG | 84.02% | 347 / 413 |
| VWB-AG | 84.47% | 87 / 103 |
- Average accuracy: 82.50%
- Weighted average accuracy: 79.36%
- Total correct: 5468 / 6890
- Bad format: 0