Datasets:
Dataset Viewer
video_id
string | step
int32 | system
string | user
string | assistant
string | image
image |
|---|---|---|---|---|---|
video_001
| 0
|
You are an expert GUI agent. You will be given a task, the history of previous actions, and two screenshots ('Before' and 'After'). Your goal is to determine the single user action that occurred between the two screenshots to progress the task.
CRITICAL: You MUST ONLY use these EXACT function names and signatures:
<code>
def click(x: float, y: float, description: str) -> str:
'''Clicks a UI element at normalized coordinates (0.0-1.0).
x, y: Normalized coordinates where 0.0 is top/left and 1.0 is bottom/right
description: Visible text on the element or its role (e.g., "Search button", "Email input field")
Example: click(0.5, 0.3, "Search button")
'''
def type(text: str) -> str:
'''Types the COMPLETE text that was entered in this action.
text: The EXACT full text visible in the After screenshot (not partial, not abbreviated)
Example: type("I want to hire a CEO for a Generative AI company in US")
'''
def scroll(direction: str, amount: str) -> str:
'''Scrolls the page or a scrollable element.
direction: Either "up" or "down"
amount: Either "small" (one scroll wheel click), "medium" (page), or "large" (multiple pages)
Example: scroll("down", "medium")
'''
</code>
ABSOLUTE REQUIREMENTS:
1. **Function Names**: ONLY use click(), type(), scroll() - NO other function names allowed
2. **No Invention**: DO NOT create functions like user_action(), press(), input(), submit(), etc.
3. **Complete Text**: For type(), extract the COMPLETE final text visible in After screenshot
4. **Real Coordinates**: For click(), provide actual normalized coordinates (0.0-1.0), not placeholders
5. **One Action Only**: Return ONE action per frame pair - the primary action that caused the change
6. **Reactive Changes**: If After shows automatic UI responses (dropdowns opening, pages loading), infer the user action that triggered it
COMMON MISTAKES TO AVOID:
❌ type("partial text") when full text is "partial text in the complete sentence"
❌ click(0.5, 0.5, "button") for EVERY click - vary coordinates based on actual position
❌ Inventing action_name() or custom_function() - stick to the three allowed functions
❌ Describing what happened instead of inferring the action
|
Please generate the next move according to the UI screenshot and expected workflow.
Expected Workflow:
Complete the following workflow:
Step 1: open_url(url="https://www.linkedin.com/")
Step 2: click(element_description="Start a post button")
Step 3: type_text(element_description="post text area", text="Are you passionate about pushing the boundaries of technology? Join us as an AI Engineer and become a key player in crafting innovative AI solutions that empower industries worldwide. At Dummy Company, we pride ourselves on fostering a culture of creativity and collaboration, and we're looking for a talented individual to create a brighter future with us. What We're Looking For; - Stay ahead of industry trends by researching and implementing state-of-the-art AI methodologies. - Strong programming skills in Python, TensorFlow, or similar frameworks. - Innovative mindset with a keen eye for detail and problem-solving. Why Join Us? - Work in a dynamic environment with opportunities for professional growth. - Enjoy a flexible work-life balance and comprehensive benefits package. - Be part of a team that values diversity, innovation, and continuous learning. Interested? Email us at [email protected] or comment below!")
Step 4: click(element_description="Post button")
Step 5: open_url(url="https://www.gmail.com/")
Step 6: click(element_description="Compose button")
Step 7: type_text(element_description="recipient field", text="[email protected]")
Step 8: type_text(element_description="subject field", text="AI Engineer at Dummy Company")
Step 9: paste_text(element_description="email body", content="Dear Candidate, Thank you for expressing interest in the AI Engineer position at Dummy Company. We are excited about the possibility of you joining our team. At Dummy Company, we pride ourselves on pushing the boundaries of technology. We have reviewed your resume and this role could be a significant step in your career. We believe that your skills and experience align perfectly with what we are looking for. Please consider applying formally to our website or reply to this email to schedule a chat with our team. We would be thrilled to further explore your potential fit for this role. Thank you once again for considering a career with us. We look forward to connecting with you soon. Warm regards, The Hiring Team Dummy Company")
Step 10: click(element_description="Send button")
Previous actions completed:
None
|
<think>
Step 2: Start a post button (from DOM ground truth with perfect frame alignment)
</think>
<code>
click(0.036, 0.014, " 0a53ac55")
</code>
| |
video_001
| 1
|
You are an expert GUI agent. You will be given a task, the history of previous actions, and two screenshots ('Before' and 'After'). Your goal is to determine the single user action that occurred between the two screenshots to progress the task.
CRITICAL: You MUST ONLY use these EXACT function names and signatures:
<code>
def click(x: float, y: float, description: str) -> str:
'''Clicks a UI element at normalized coordinates (0.0-1.0).
x, y: Normalized coordinates where 0.0 is top/left and 1.0 is bottom/right
description: Visible text on the element or its role (e.g., "Search button", "Email input field")
Example: click(0.5, 0.3, "Search button")
'''
def type(text: str) -> str:
'''Types the COMPLETE text that was entered in this action.
text: The EXACT full text visible in the After screenshot (not partial, not abbreviated)
Example: type("I want to hire a CEO for a Generative AI company in US")
'''
def scroll(direction: str, amount: str) -> str:
'''Scrolls the page or a scrollable element.
direction: Either "up" or "down"
amount: Either "small" (one scroll wheel click), "medium" (page), or "large" (multiple pages)
Example: scroll("down", "medium")
'''
</code>
ABSOLUTE REQUIREMENTS:
1. **Function Names**: ONLY use click(), type(), scroll() - NO other function names allowed
2. **No Invention**: DO NOT create functions like user_action(), press(), input(), submit(), etc.
3. **Complete Text**: For type(), extract the COMPLETE final text visible in After screenshot
4. **Real Coordinates**: For click(), provide actual normalized coordinates (0.0-1.0), not placeholders
5. **One Action Only**: Return ONE action per frame pair - the primary action that caused the change
6. **Reactive Changes**: If After shows automatic UI responses (dropdowns opening, pages loading), infer the user action that triggered it
COMMON MISTAKES TO AVOID:
❌ type("partial text") when full text is "partial text in the complete sentence"
❌ click(0.5, 0.5, "button") for EVERY click - vary coordinates based on actual position
❌ Inventing action_name() or custom_function() - stick to the three allowed functions
❌ Describing what happened instead of inferring the action
|
Please generate the next move according to the UI screenshot and expected workflow.
Expected Workflow:
Complete the following workflow:
Step 1: open_url(url="https://www.linkedin.com/")
Step 2: click(element_description="Start a post button")
Step 3: type_text(element_description="post text area", text="Are you passionate about pushing the boundaries of technology? Join us as an AI Engineer and become a key player in crafting innovative AI solutions that empower industries worldwide. At Dummy Company, we pride ourselves on fostering a culture of creativity and collaboration, and we're looking for a talented individual to create a brighter future with us. What We're Looking For; - Stay ahead of industry trends by researching and implementing state-of-the-art AI methodologies. - Strong programming skills in Python, TensorFlow, or similar frameworks. - Innovative mindset with a keen eye for detail and problem-solving. Why Join Us? - Work in a dynamic environment with opportunities for professional growth. - Enjoy a flexible work-life balance and comprehensive benefits package. - Be part of a team that values diversity, innovation, and continuous learning. Interested? Email us at [email protected] or comment below!")
Step 4: click(element_description="Post button")
Step 5: open_url(url="https://www.gmail.com/")
Step 6: click(element_description="Compose button")
Step 7: type_text(element_description="recipient field", text="[email protected]")
Step 8: type_text(element_description="subject field", text="AI Engineer at Dummy Company")
Step 9: paste_text(element_description="email body", content="Dear Candidate, Thank you for expressing interest in the AI Engineer position at Dummy Company. We are excited about the possibility of you joining our team. At Dummy Company, we pride ourselves on pushing the boundaries of technology. We have reviewed your resume and this role could be a significant step in your career. We believe that your skills and experience align perfectly with what we are looking for. Please consider applying formally to our website or reply to this email to schedule a chat with our team. We would be thrilled to further explore your potential fit for this role. Thank you once again for considering a career with us. We look forward to connecting with you soon. Warm regards, The Hiring Team Dummy Company")
Step 10: click(element_description="Send button")
Previous actions completed:
Step 1: click(0.036, 0.014, " 0a53ac55")
|
<think>
Step 4: Post button (from DOM ground truth with perfect frame alignment)
</think>
<code>
click(0.231, 0.072, "aria")
</code>
|
UI Automation Dataset (Multi-Video)
2 examples from 1 videos - UI automation tasks from screen recordings.
Dataset Structure
Each entry contains:
- video_id: Sequential ID for each video (video_001, video_002, etc.)
- step: Step number within that video (0, 1, 2, ...)
- system: System prompt for the GUI agent
- user: Task instruction + previous actions
- assistant: Model's reasoning and action
- image: Screenshot of the UI state
Usage
from datasets import load_dataset
ds = load_dataset("KMH158-QLU/recruiter_perfect2")
# Access by video
for video_id in set(ds['train']['video_id']):
video_data = ds['train'].filter(lambda x: x['video_id'] == video_id)
print(f"Video {video_id}: {len(video_data)} steps")
# Or iterate all examples
for item in ds['train']:
print(f"{item['video_id']} - Step {item['step']}: {item['assistant'][:50]}...")
Growing Dataset
This dataset supports multiple videos. Each video gets a unique ID (video_001, video_002, etc.). New videos are automatically appended with the next available ID.
- Downloads last month
- 16