AppGen Qwen3 Continue-SFT Norm1000 Navigation Thinking
Continue-SFT checkpoint of luca0621/appgen-qwen3-thinktag-lr5e-7-ep1 trained on AMEX-derived navigation data converted to normalized mobile_use tool calls with an empty Qwen3 think block before the tool call.
- Checkpoint:
checkpoint-2406 - Training output:
continue_sft_qwen3_mobile_use_norm1000_nav_only_thinking_gpu23_20260625_r2 - Format:
mobile_use_norm1000 - Prompt/eval style:
gelab - Training: 1 epoch, lr
5e-7, full finetune, visual frozen, DeepSpeed ZeRO-2, GPUs 2-3 - W&B run: https://wandb.ai/namhokoh-korea-advanced-institute-of-science-and-technology/appgen-continue-train/runs/z2d17khd
Real-World GUI Eval
Evaluated with eval/evaluate_real_world_aligned.py on the real-world benchmark suite.
| Benchmark | Accuracy | Correct / Total | BadFmt |
|---|---|---|---|
| ScreenSpot | 90.25% | 1148 / 1272 | 0 |
| ScreenSpot-v2 | 93.16% | 1185 / 1272 | 0 |
| FuncPred | 62.05% | 1241 / 2000 | 0 |
| MoTIF | 76.52% | 968 / 1265 | 0 |
| Refexp | 85.84% | 485 / 565 | 0 |
| VWB-EG | 84.02% | 347 / 413 | 0 |
| VWB-AG | 84.47% | 87 / 103 | 0 |
Average: 82.33%
Weighted average: 79.26% (5461 / 6890)