菜单
Enable training and evaluation of desktop computer-using agents through 1.2M action steps across diverse Windows applications, covering GUI grounding, screen parsing, and action prediction with hybrid GUI+API action space reflecting modern agent designs.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill gui-360-desktop-agent-dataset命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
基于 SOC 职业分类
| name | gui-360-desktop-agent-dataset |
| title | GUI-360°: Comprehensive Dataset for Computer-Using Agents |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2511.04307 |
| keywords | ["Desktop Agents","GUI Grounding","Action Prediction","Multimodal Agents","Autonomous Agents"] |
| description | Enable training and evaluation of desktop computer-using agents through 1.2M action steps across diverse Windows applications, covering GUI grounding, screen parsing, and action prediction with hybrid GUI+API action space reflecting modern agent designs. |
Desktop agents that autonomously complete computer tasks face unique challenges: grounding spatial coordinates, parsing complex UIs, predicting appropriate next actions. GUI-360° provides 1.2M executed action steps across thousands of real trajectories in Word, Excel, and PowerPoint, with full screenshots, accessibility metadata, task goals, reasoning traces, and both successful and failed action sequences. The dataset reveals that state-of-the-art models struggle with grounding and action prediction despite strong general capabilities.
The benchmark spans three canonical tasks: GUI grounding, screen parsing, and action prediction.
Comprehensive Desktop Agent Training and Evaluation:
1. Design Data Collection Pipeline
Capture human demonstrations with full metadata.
class DesktopAgentDataCollector:
def __init__(self, recording_dir="./gui_recordings"):
self.recording_dir = recording_dir
self.current_trajectory =
():
.current_trajectory = {
: task_description,
: [],
: [],
: {
: ,
: datetime.now(),
: []
}
}
():
screenshot = .screenshot_full_screen()
a11y_tree = .get_accessibility_tree()
step = {
: {
: action_type,
: coordinates,
: text,
: keystroke
},
: screenshot,
: a11y_tree,
: datetime.now()
}
.current_trajectory[].append(step)
(.current_trajectory[])
():
.current_trajectory[][] = task_completed
.current_trajectory[][] = final_notes
.current_trajectory[][] = datetime.now()
trajectory_id = .save_trajectory(.current_trajectory)
trajectory_id
():
pyautogui
screenshot = pyautogui.screenshot()
np.array(screenshot)
():
pywinauto
app = pywinauto.GetFocusedWindow()
a11y_tree = .build_tree(app)
a11y_tree
():
current_depth > max_depth:
tree = {
: element.element_info.name,
: element.element_info.control_type,
: element.rectangle().to_dict(),
: element.is_enabled(),
: element.is_visible(),
: []
}
:
child element.children():
child_tree = .build_tree(child, max_depth, current_depth + )
child_tree:
tree[].append(child_tree)
:
tree
2. Implement Three Evaluation Tasks
Structure benchmark around distinct technical challenges.
class GUIAgentBenchmark:
def __init__(self, dataset_path):
self.trajectories = self.load_trajectories(dataset_path)
def task1_gui_grounding(self, model):
"""Predict screen coordinates for target actions"""
results = []
for trajectory in self.trajectories:
for step in trajectory['steps']:
# Input: screenshot + action description
# Output: coordinate prediction
screenshot = step['screenshot']
action_desc = step['action']['type']
if step['action']['type'] == 'click':
true_coords = step['action']['coordinates']
# Model predicts coordinate
pred_coords = model.predict_click_location(screenshot, action_desc)
# Distance metric
distance = np.linalg.norm(np.array(pred_coords) - np.array(true_coords))
results.append({
'task': 'grounding',
'distance_error': distance,
'correct': distance < 50 # Within 50 pixels
})
accuracy = np.mean([r['correct'] for r in results])
return {'accuracy': accuracy, 'results': results}
def task2_screen_parsing(self, model):
"""Predict UI elements and their properties"""
results = []
for trajectory in self.trajectories:
for step in trajectory['steps']:
screenshot = step['screenshot']
true_a11y_tree = step['accessibility_tree']
# Model parses screen
pred_elements = model.parse_screen(screenshot)
# Match predictions to ground truth
# Metrics: precision, recall, F1 on element detection
metrics = self.compute_parsing_metrics(pred_elements, true_a11y_tree)
results.append(metrics)
avg_f1 = np.mean([r['f1'] for r in results])
return {'f1': avg_f1, 'results': results}
def task3_action_prediction(self, model):
"""Predict next action given current state and goal"""
results = []
for trajectory in self.trajectories:
for i in range(len(trajectory['steps']) - 1):
current_step = trajectory['steps'][i]
next_step = trajectory['steps'][i + 1]
# Input: current screenshot + goal description
screenshot = current_step['screenshot']
goal = trajectory['task']
# Ground truth next action
true_action = next_step['action']
# Model predicts action
pred_action = model.predict_next_action(screenshot, goal)
# Accuracy: action type matches
action_correct = pred_action['type'] == true_action['type']
results.append({
'action_correct': action_correct,
'action_type': true_action['type']
})
accuracy = np.mean([r['action_correct'] for r in results])
accuracy_by_type = {}
for action_type in set(r['action_type'] for r in results):
type_results = [r for r in results if r['action_type'] == action_type]
accuracy_by_type[action_type] = np.mean([r['action_correct'] for r in type_results])
return {'accuracy': accuracy, 'by_type': accuracy_by_type, 'results': results}
def compute_parsing_metrics(self, predicted, ground_truth):
"""Compare parsed UI elements to ground truth"""
# Extract ground truth elements
gt_elements = self.extract_elements(ground_truth)
# Precision: predicted elements that match ground truth
true_positives = 0
for pred in predicted:
for gt in gt_elements:
if self.elements_match(pred, gt):
true_positives += 1
break
precision = true_positives / len(predicted) if predicted else 0
recall = true_positives / len(gt_elements) if gt_elements else 0
f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0
return {'precision': precision, 'recall': recall, 'f1': f1}
3. Implement Hybrid Action Execution
Support both GUI and API-based actions.
class HybridActionExecutor:
def __init__(self, app_type='word'):
self.app_type = app_type
self.gui_executor = GUIActionExecutor()
self.api_executor = APIActionExecutor(app_type)
def execute_action(self, action_spec):
"""Execute action using appropriate method"""
action_type = action_spec['type']
if action_type in ['click', 'type', 'drag', 'scroll']:
# GUI action
return self.gui_executor.execute(action_spec)
elif action_type in ['table_insert', 'cell_set', 'slide_modify']:
# API action (app-specific)
return self.api_executor.execute(action_spec)
else:
raise ValueError(f"Unknown action type: {action_type}")
class GUIActionExecutor:
def execute(self, action_spec):
"""Execute GUI action via mouse/keyboard"""
import pyautogui
action_type = action_spec['type']
if action_type == 'click':
pyautogui.click(action_spec['coordinates'])
elif action_type == 'type':
pyautogui.typewrite(action_spec['text'], interval=0.05)
elif action_type == 'drag':
start = action_spec['start_coords']
end = action_spec['end_coords']
pyautogui.moveTo(start)
pyautogui.drag(end[0] - start[0], end[1] - start[1], duration=0.5)
elif action_type == 'scroll':
amount = action_spec['amount']
pyautogui.scroll(amount)
return {'success': True}
class APIActionExecutor:
def __init__(self, app_type):
self.app_type = app_type
self.excel_api = ExcelAPI() if app_type == 'excel' else None
self.word_api = WordAPI() if app_type == 'word' else None
def execute(self, action_spec):
"""Execute action via application API"""
if self.app_type == 'excel':
return self.excel_api.execute(action_spec)
elif self.app_type == 'word':
return self.word_api.execute(action_spec)
class ExcelAPI:
def execute(self, action_spec):
"""Execute Excel-specific actions"""
from openpyxl import load_workbook
action_type = action_spec['type']
if action_type == 'cell_set':
wb = load_workbook(action_spec['file'])
ws = wb[action_spec['sheet']]
ws[action_spec['cell']] = action_spec['value']
wb.save(action_spec['file'])
elif action_type == 'auto_fill':
ws = self.get_worksheet(action_spec)
start_cell = action_spec['start']
fill_range = action_spec['range']
# Implementation details...
return {'success': True}
When to Use:
Hyperparameters:
When NOT to Use:
Pitfalls:
Integration Strategy: Use as pre-training data for agents. Fine-tune on domain-specific applications after. Combine with synthetic data augmentation for scaling.