stephenfan1101/qwen25vl7b_baseline_checkpoint-4000

Qwen2.5-VL-7B coordinate-text baseline, checkpoint-4000

ScreenSpot-Pro hit_top1 All-avg = 21.57 (n=1581).

OOM 前最后一个完整 checkpoint。曲线在 2000 之后回落，尚未跑过 step 6000 的预定止损点。

What this run is

The pointer-free reference arm for the 7B SFT suite. Same data and global batch as the pointer-head arms, but the model is stock Qwen2.5-VL-7B-Instruct trained to emit pyautogui.click(x=..., y=...) in smart-resized pixels. There is no pointer head. The visual encoder is trained (unfreeze_all_parameters).

setting	value
learning rate	1e-6
weight decay	0.01
warmup	0.03
schedule	cosine
global batch	64 (4 x 8 x micro-1 x accum-2)
save every	2000 steps
planned length	19613 steps (one epoch)

Sibling pointer-head peaks on the same data sit near 46-47. This arm is the coordinate-text ceiling, not a pointer-head competitor.

Contents

Consolidated weights, tokenizer, trainer_state.json. DeepSpeed ZeRO-3 optimizer shards are omitted: they are 24G per node and only reloadable on the original 32-GPU layout.