Stevenshuqing's picture
Upload README.md with huggingface_hub
89a3556 verified
metadata
tags:
  - gui-agent
  - qwen2.5-vl
  - cooperative-lora
  - gui-360
license: apache-2.0
base_model: Qwen/Qwen2.5-VL-7B-Instruct
datasets:
  - Stevenshuqing/gui360-balanced

gui360-fullparam-sft-step250

Full-parameter SFT on GUI-360 balanced 2K data (Qwen2.5-VL-7B-Instruct). TSR=22.2%, StepSR=69.3%. Best overall. ~epoch 3.7, LLaMA-Factory+ZeRO-3.

Base Model

  • Qwen2.5-VL-7B-Instruct

Training Data

  • GUI-360 balanced 2K episodes (17,264 steps)
  • Action types: click, type, swipe (balanced sampling)

Evaluation (GUI-360 test 1K balanced)

Metric Value
TSR (Task Success Rate) 22.2%
StepSR (Step Success Rate) 69.3%
Progress 35.3%

Full Ranking

# Method TSR Params
1 Full-param SFT step-250 22.2% 7.6B
2 V15 Cooperative RL step-25 20.8% ~142M
3 PEFT Cooperative r=128 (SVD) 18.6% ~67M
4 PEFT Standard r=128 (SVD) 18.1% ~67M
5 Base model (zero-shot) 2.4%

Citation

Part of the Cooperative LoRA research for GUI agents.