Continue-SFT of luca0621/appgen-qwen3-thinktag-lr5e-7-ep1 (Qwen3-VL) on AMEX-derived GUI trajectories converted to the mobile_use tool-call format with normalized 0-1000 coordinates.

Base model: luca0621/appgen-qwen3-thinktag-lr5e-7-ep1
Checkpoint: checkpoint-2420
Format: <tool_call>{"name": "mobile_use", "arguments": {"action": "click", "coordinate": [x, y]}}</tool_call>
Coordinate space: normalized 0-1000 for screen/image coordinates
Training: 1 epoch, learning rate 5e-7, max length 4096, max pixels 2629536
W&B run: qwen3-mobile-use-norm1000-continue-sft-gpu01-rerun

Real-World GUI Grounding Eval

Point-in-box accuracy with normalized 0-1000 coordinates:

Benchmark	Accuracy	Correct / Total
ScreenSpot	90.57%	1152 / 1272
ScreenSpot-v2	93.40%	1188 / 1272
FuncPred	62.25%	1245 / 2000
MoTIF	76.92%	973 / 1265
Refexp	85.84%	485 / 565
VWB-EG	83.78%	346 / 413
VWB-AG	83.50%	86 / 103

Average accuracy: 82.32%
Weighted average accuracy: 79.46%