Complete Label Studio Tutorial: Data Labeling for Machine Learning

Label Studio is an open-source data labeling platform for machine learning. It supports labeling various data types including text, images, audio, video, and time series, making it essential for building high-quality ML training datasets.

Why Label Studio?

Label Studio Advantages:

Multi-modal: Label text, images, audio, video, HTML
Flexible: Customizable labeling interfaces
Collaborative: Team-based annotation workflows
Integrations: ML backends, cloud storage, webhooks
Open source: Self-hosted with enterprise options

Use Cases:

NLP annotation (NER, sentiment, classification)
Computer vision labeling (bounding boxes, segmentation)
Audio transcription and classification
Multi-modal data annotation
Active learning workflows

Installation

# Install with pip pip install label-studio Start Label Studio label-studio start Or with Docker docker run -it -p 8080:8080 \ -v $(pwd)/mydata:/label-studio/data \ heartexlabs/label-studio:latest Access at http://localhost:8080

Quick Start

1. Create Project

from labelstudiosdk import Client

Connect to Label Studio
ls = Client(url='http://localhost:8080', apikey='your-api-key')


Create project
project = ls.startproject(
    title='Sentiment Analysis',
    labelconfig='''

    
      
      
        
        
        
      
    
    '''
)

print(f"Project created: {project.id}")

2. Import Data

# Import from list
tasks = [
    {"text": "I love this product!"},
    {"text": "This is terrible."},
    {"text": "It's okay, nothing special."}
]

project.importtasks(tasks)

Import from file
project.importtasks('data.json')


Import from URL
project.importtasks([
    {"image": "https://example.com/image1.jpg"},
    {"image": "https://example.com/image2.jpg"}
])

3. Export Annotations

# Export all annotations
annotations = project.exporttasks()


Export in specific format
annotations = project.exporttasks(exporttype='JSON')

annotations = project.exporttasks(exporttype='CSV')

annotations = project.exporttasks(exporttype='COCO')

annotations = project.exporttasks(exporttype='YOLO')


Save to file
import json
with open('annotations.json', 'w') as f:
    json.dump(annotations, f)

Label Configurations

1. Text Classification

2. Named Entity Recognition

3. Image Classification

4. Object Detection (Bounding Boxes)

5. Image Segmentation

6. Audio Transcription

<p> rows="4" editable="true" maxSubmissions="1"/></p> <p></View></p> <p></code></pre></p> <h3 id="7-audio-classification-with-regions">7. Audio Classification with Regions</h3> <pre><code><View> <p> <Labels name="label" toName="audio"></p> <p> <Label value="Speech"/></p> <p> <Label value="Music"/></p> <p> <Label value="Noise"/></p> <p> </Labels></p> <p> <Audio name="audio" value="$audio"/></p> <p></View></p> <p></code></pre></p> <h3 id="8-video-annotation">8. Video Annotation</h3> <pre><code><View> <p> <Video name="video" value="$video"/></p> <p> <VideoRectangle name="box" toName="video"/></p> <p> <Labels name="label" toName="video"></p> <p> <Label value="Person"/></p> <p> <Label value="Vehicle"/></p> <p> </Labels></p> <p></View></p> <p></code></pre></p> <h2 id="sdk-operations">SDK Operations</h2> <h3 id="1-project-management">1. Project Management</h3> <pre><code>from label</em>studio<em>sdk import Client <p>ls = Client(url='http://localhost:8080', api</em>key='your-api-key')</p> <h1>List all projects</h1> <p>projects = ls.get<em>projects()</p> <p>for p in projects:</p> <p> print(f"{p.id}: {p.title}")</p> <h1>Get specific project</h1> <p>project = ls.get</em>project(project<em>id=1)</p> <h1>Update project</h1> <p>project.set</em>params(</p> <p> title="Updated Title",</p> <p> description="New description"</p> <p>)</p> <h1>Delete project</h1> <p>project.delete()</p> <p></code></pre></p> <h3 id="2-task-management">2. Task Management</h3> <pre><code># Get all tasks <p>tasks = project.get<em>tasks()</p> <h1>Get specific task</h1> <p>task = project.get</em>task(task<em>id=1)</p> <h1>Update task</h1> <p>project.update</em>task(</p> <p> task<em>id=1,</p> <p> data={"text": "Updated text"}</p> <p>)</p> <h1>Delete task</h1> <p>project.delete</em>task(task<em>id=1)</p> <h1>Get tasks with filters</h1> <p>from label</em>studio<em>sdk.data</em>manager import Filters, Column, Operator, Type</p> <p>filters = Filters.create(Filters.AND, [</p> <p> Filters.item(</p> <p> Column.completed<em>at,</p> <p> Operator.EMPTY,</p> <p> Type.Datetime,</p> <p> Filters.value(True)</p> <p> )</p> <p>])</p> <p>unlabeled</em>tasks = project.get<em>tasks(filters=filters)</p> <p></code></pre></p> <h3 id="3-annotation-management">3. Annotation Management</h3> <pre><code># Create annotation <p>annotation = project.create</em>annotation(</p> <p> task<em>id=1,</p> <p> result=[</p> <p> {</p> <p> "from</em>name": "sentiment",</p> <p> "to<em>name": "text",</p> <p> "type": "choices",</p> <p> "value": {"choices": ["Positive"]}</p> <p> }</p> <p> ]</p> <p>)</p> <h1>Get annotations for task</h1> <p>annotations = project.get</em>annotations(task<em>id=1)</p> <h1>Delete annotation</h1> <p>project.delete</em>annotation(annotation<em>id=1)</p> <p></code></pre></p> <h3 id="4-prediction-upload">4. Prediction Upload</h3> <pre><code># Upload predictions for pre-labeling <p>predictions = [</p> <p> {</p> <p> "task": 1,</p> <p> "result": [</p> <p> {</p> <p> "from</em>name": "sentiment",</p> <p> "to<em>name": "text",</p> <p> "type": "choices",</p> <p> "value": {"choices": ["Positive"]}</p> <p> }</p> <p> ],</p> <p> "score": 0.95</p> <p> }</p> <p>]</p> <p>project.create</em>predictions(predictions)</p> <h1>Bulk prediction upload</h1> <p>model<em>predictions = []</p> <p>for task in project.get</em>tasks():</p> <p> pred = model.predict(task['data']['text'])</p> <p> model<em>predictions.append({</p> <p> "task": task['id'],</p> <p> "result": [{</p> <p> "from</em>name": "sentiment",</p> <p> "to<em>name": "text",</p> <p> "type": "choices",</p> <p> "value": {"choices": [pred['label']]}</p> <p> }],</p> <p> "score": pred['confidence']</p> <p> })</p> <p>project.create</em>predictions(model<em>predictions)</p> <p></code></pre></p> <h2 id="cloud-storage-integration">Cloud Storage Integration</h2> <h3 id="1-amazon-s3">1. Amazon S3</h3> <pre><code># Connect S3 storage <p>storage = project.connect</em>s3<em>import</em>storage(</p> <p> bucket='my-bucket',</p> <p> prefix='images/',</p> <p> aws<em>access</em>key<em>id='AKIAXXXXXXXX',</p> <p> aws</em>secret<em>access</em>key='xxxxxxxxxx',</p> <p> region<em>name='us-east-1',</p> <p> recursive</em>scan=True</p> <p>)</p> <h1>Sync storage</h1> <p>storage.sync()</p> <h1>Export to S3</h1> <p>export<em>storage = project.connect</em>s3<em>export</em>storage(</p> <p> bucket='my-bucket',</p> <p> prefix='annotations/',</p> <p> aws<em>access</em>key<em>id='AKIAXXXXXXXX',</p> <p> aws</em>secret<em>access</em>key='xxxxxxxxxx'</p> <p>)</p> <p></code></pre></p> <h3 id="2-google-cloud-storage">2. Google Cloud Storage</h3> <pre><code># Connect GCS storage <p>storage = project.connect<em>google</em>import<em>storage(</p> <p> bucket='my-bucket',</p> <p> prefix='data/',</p> <p> google</em>application<em>credentials='path/to/credentials.json'</p> <p>)</p> <p>storage.sync()</p> <p></code></pre></p> <h3 id="3-azure-blob-storage">3. Azure Blob Storage</h3> <pre><code># Connect Azure storage <p>storage = project.connect</em>azure<em>import</em>storage(</p> <p> container='my-container',</p> <p> prefix='images/',</p> <p> account<em>name='myaccount',</p> <p> account</em>key='xxxxxxxxxx'</p> <p>)</p> <p>storage.sync()</p> <p></code></pre></p> <h2 id="ml-backend-integration">ML Backend Integration</h2> <h3 id="1-create-ml-backend">1. Create ML Backend</h3> <pre><code># ml<em>backend.py <p>from label</em>studio<em>ml import LabelStudioMLBase</p> <p>class SentimentModel(LabelStudioMLBase):</p> <p> def <strong>init</strong>(self, <strong>kwargs):</p> <p> super(SentimentModel, self).<strong>init</strong>(</strong>kwargs)</p> <p> # Load your model</p> <p> from transformers import pipeline</p> <p> self.model = pipeline("sentiment-analysis")</p> <p> def predict(self, tasks, <strong>kwargs):</p> <p> predictions = []</p> <p> for task in tasks:</p> <p> text = task['data']['text']</p> <p> result = self.model(text)[0]</p> <p> label = "Positive" if result['label'] == 'POSITIVE' else "Negative"</p> <p> predictions.append({</p> <p> 'result': [{</p> <p> 'from</em>name': 'sentiment',</p> <p> 'to<em>name': 'text',</p> <p> 'type': 'choices',</p> <p> 'value': {'choices': [label]}</p> <p> }],</p> <p> 'score': result['score']</p> <p> })</p> <p> return predictions</p> <p> def fit(self, annotations, </strong>kwargs):</p> <p> # Train your model with new annotations</p> <p> pass</p> <p></code></pre></p> <h3 id="2-run-ml-backend">2. Run ML Backend</h3> <pre><code># Install ML backend <p>pip install label-studio-ml</p> <h1>Initialize backend</h1> <p>label-studio-ml init my</em>backend --script ml<em>backend.py</p> <h1>Start backend</h1> <p>label-studio-ml start my</em>backend -p 9090</p> <p></code></pre></p> <h3 id="3-connect-to-label-studio">3. Connect to Label Studio</h3> <pre><code># Add ML backend to project <p>project.connect<em>ml</em>backend(</p> <p> url='http://localhost:9090',</p> <p> title='Sentiment Model'</p> <p>)</p> <h1>Trigger predictions</h1> <p>project.start<em>predictions()</p> <p></code></pre></p> <h2 id="active-learning">Active Learning</h2> <h3 id="1-setup-active-learning">1. Setup Active Learning</h3> <pre><code>from label</em>studio<em>ml import LabelStudioMLBase <p>import numpy as np</p> <p>class ActiveLearningModel(LabelStudioMLBase):</p> <p> def predict(self, tasks, <strong>kwargs):</p> <p> predictions = []</p> <p> for task in tasks:</p> <p> # Get model prediction with uncertainty</p> <p> pred, confidence = self.model.predict</em>with<em>uncertainty(task['data'])</p> <p> predictions.append({</p> <p> 'result': [{</p> <p> 'from</em>name': 'label',</p> <p> 'to<em>name': 'text',</p> <p> 'type': 'choices',</p> <p> 'value': {'choices': [pred]}</p> <p> }],</p> <p> 'score': confidence</p> <p> })</p> <p> return predictions</p> <p> def get</em>tasks<em>to</em>label(self, n=10):</p> <p> # Get tasks sorted by uncertainty</p> <p> tasks = self.project.get<em>tasks()</p> <p> predictions = self.predict(tasks)</p> <p> # Sort by uncertainty (lowest confidence first)</p> <p> sorted</em>tasks = sorted(</p> <p> zip(tasks, predictions),</p> <p> key=lambda x: x[1]['score']</p> <p> )</p> <p> return [t[0] for t in sorted<em>tasks[:n]]</p> <p></code></pre></p> <h3 id="2-uncertainty-sampling">2. Uncertainty Sampling</h3> <pre><code>from label</em>studio<em>sdk import Client <p>ls = Client(url='http://localhost:8080', api</em>key='your-api-key')</p> <p>project = ls.get<em>project(project</em>id=1)</p> <h1>Get tasks with prediction scores</h1> <p>tasks = project.get<em>tasks()</p> <h1>Sort by prediction score (uncertainty)</h1> <p>uncertain</em>tasks = sorted(</p> <p> [t for t in tasks if t.get('predictions')],</p> <p> key=lambda t: min(p['score'] for p in t['predictions'])</p> <p>)</p> <h1>Prioritize uncertain tasks</h1> <p>for task in uncertain<em>tasks[:10]:</p> <p> print(f"Task {task['id']}: Score {task['predictions'][0]['score']}")</p> <p></code></pre></p> <h2 id="webhooks">Webhooks</h2> <h3 id="1-setup-webhooks">1. Setup Webhooks</h3> <pre><code># Create webhook <p>webhook = project.create</em>webhook(</p> <p> url='https://my-server.com/webhook',</p> <p> send<em>payload=True,</p> <p> send</em>for<em>all</em>actions=False,</p> <p> actions=['ANNOTATION<em>CREATED', 'ANNOTATION</em>UPDATED']</p> <p>)</p> <h1>Available actions:</h1> <h1>- ANNOTATION<em>CREATED</h1> <h1>- ANNOTATION</em>UPDATED</h1> <h1>- ANNOTATION<em>DELETED</h1> <h1>- TASK</em>CREATED</h1> <h1>- PROJECT<em>UPDATED</h1> <p></code></pre></p> <h3 id="2-handle-webhook">2. Handle Webhook</h3> <pre><code>from flask import Flask, request <p>app = Flask(<strong>name</strong>)</p> <p>@app.route('/webhook', methods=['POST'])</p> <p>def handle</em>webhook():</p> <p> data = request.json</p> <p> action = data['action']</p> <p> annotation = data.get('annotation')</p> <p> task = data.get('task')</p> <p> if action == 'ANNOTATION<em>CREATED':</p> <p> # Trigger model retraining</p> <p> print(f"New annotation for task {task['id']}")</p> <p> # retrain</em>model(annotation)</p> <p> return {'status': 'ok'}</p> <p>if <strong>name</strong> == '<strong>main</strong>':</p> <p> app.run(port=5000)</p> <p></code></pre></p> <h2 id="quality-control">Quality Control</h2> <h3 id="1-agreement-metrics">1. Agreement Metrics</h3> <pre><code># Get agreement scores <p>agreements = project.get<em>agreement()</p> <p>for task</em>id, agreement in agreements.items():</p> <p> print(f"Task {task<em>id}: {agreement['agreement']:.2%}")</p> <p></code></pre></p> <h3 id="2-review-workflow">2. Review Workflow</h3> <pre><code><View> <p> <Text name="text" value="$text"/></p> <p> <Choices name="sentiment" toName="text"></p> <p> <Choice value="Positive"/></p> <p> <Choice value="Negative"/></p> <p> </Choices></p> <p> </p> <p> <Header value="Review"/></p> <p> <Choices name="review" toName="text" choice="single"></p> <p> <Choice value="Approved"/></p> <p> <Choice value="Rejected"/></p> <p> <Choice value="Needs Review"/></p> <p> </Choices></p> <p> <TextArea name="review</em>comment" toName="text" placeholder="Review comments"/></p> <p></View></p> <p></code></pre></p> <h2 id="best-practices">Best Practices</h2> <h3 id="1-efficient-data-import">1. Efficient Data Import</h3> <pre><code>import json <p>from label<em>studio</em>sdk import Client</p> <p>ls = Client(url='http://localhost:8080', api<em>key='your-api-key')</p> <p>project = ls.get</em>project(project<em>id=1)</p> <h1>Batch import for large datasets</h1> <p>def batch</em>import(data, batch<em>size=1000):</p> <p> for i in range(0, len(data), batch</em>size):</p> <p> batch = data[i:i+batch<em>size]</p> <p> project.import</em>tasks(batch)</p> <p> print(f"Imported {min(i+batch<em>size, len(data))}/{len(data)}")</p> <h1>Load and import</h1> <p>with open('large</em>dataset.json') as f:</p> <p> data = json.load(f)</p> <p>batch<em>import(data)</p> <p></code></pre></p> <h3 id="2-export-processing">2. Export Processing</h3> <pre><code>import pandas as pd <h1>Export annotations</h1> <p>annotations = project.export</em>tasks()</p> <h1>Process for ML training</h1> <p>training<em>data = []</p> <p>for task in annotations:</p> <p> if task['annotations']:</p> <p> annotation = task['annotations'][0]</p> <p> for result in annotation['result']:</p> <p> if result['type'] == 'choices':</p> <p> training</em>data.append({</p> <p> 'text': task['data']['text'],</p> <p> 'label': result['value']['choices'][0]</p> <p> })</p> <p>df = pd.DataFrame(training<em>data)</p> <p>df.to</em>csv('training_data.csv', index=False)</p> <p></code></pre></p> <h2 id="conclusion">Conclusion</h2> <p>Label Studio is essential for ML data labeling with:</p> <li></strong>Multi-modal support<strong>: Text, image, audio, video</li> <li></strong>Flexible interfaces<strong>: Customizable labeling configs</li> <li></strong>Collaboration<strong>: Team annotation workflows</li> <li></strong>ML integration<strong>: Backend predictions, active learning</li> <li></strong>Cloud storage**: S3, GCS, Azure integration</li> <p>Key takeaways:</p> <ul><li>Design clear labeling interfaces</li> <li>Use predictions for pre-labeling efficiency</li> <li>Implement quality control workflows</li> <li>Leverage webhooks for automation</li> <li>Batch operations for large datasets</li> <p></ul></p> </div> <footer style="margin-top:2rem;padding-top:1rem;border-top:1px solid #2a2c35;font-size:0.875rem;color:#9ca0ab;"> <p>Written by Ruby Abdullah at <a href="https://rubythalib.ai">rubythalib.ai</a></p> </footer> </article> <section style="max-width:800px;margin:3rem auto 0;padding:2rem 1rem 0;background:#0b0b0f;color:#e6e7eb;border-top:1px solid #2a2c35;"> <h2 style="font-size:1.5rem;font-weight:600;margin-bottom:1rem;">Related Articles</h2> <div style="display:grid;grid-template-columns:repeat(auto-fit,minmax(220px,1fr));gap:1rem;"> <a href="/en/articles/tutorial-lengkap-comet-ml-platform-mlops-untuk-experiment-tracking-dan-model-management" style="text-decoration:none;color:inherit;display:block;"> <div style="border:1px solid #2a2c35;border-radius:8px;padding:1rem;height:100%;background:#15161c;"> <h3 style="font-size:1rem;font-weight:600;margin:0 0 0.5rem 0;">Complete Comet ML Tutorial: MLOps Platform for Experiment Tracking and Model Management</h3> <p style="font-size:0.875rem;color:#9ca0ab;margin:0;">Tutorial Lengkap Comet ML: Platform MLOps untuk Experiment Tracking dan Model Management Dalam dunia machine learning mo...</p> </div> </a> <a href="/en/articles/tutorial-surya-ocr-document-ai-untuk-text-detection-ocr-dan-layout-analysis" style="text-decoration:none;color:inherit;display:block;"> <div style="border:1px solid #2a2c35;border-radius:8px;padding:1rem;height:100%;background:#15161c;"> <h3 style="font-size:1rem;font-weight:600;margin:0 0 0.5rem 0;">Surya OCR Tutorial: Document AI for Text Detection, OCR, and Layout Analysis</h3> <p style="font-size:0.875rem;color:#9ca0ab;margin:0;">Tutorial Surya OCR: Document AI untuk Text Detection, OCR, dan Layout Analysis Surya adalah library Document AI open-sou...</p> </div> </a> <a href="/en/articles/tutorial-flyte-workflow-orchestration-untuk-machine-learning-dan-data-engineering" style="text-decoration:none;color:inherit;display:block;"> <div style="border:1px solid #2a2c35;border-radius:8px;padding:1rem;height:100%;background:#15161c;"> <h3 style="font-size:1rem;font-weight:600;margin:0 0 0.5rem 0;">Flyte Tutorial: Workflow Orchestration for Machine Learning and Data Engineering</h3> <p style="font-size:0.875rem;color:#9ca0ab;margin:0;">Tutorial Flyte: Workflow Orchestration untuk Machine Learning dan Data Engineering Flyte adalah platform workflow orches...</p> </div> </a> <a href="/en/articles/tutorial-litserve-framework-serving-model-ai-yang-cepat-dan-mudah" style="text-decoration:none;color:inherit;display:block;"> <div style="border:1px solid #2a2c35;border-radius:8px;padding:1rem;height:100%;background:#15161c;"> <h3 style="font-size:1rem;font-weight:600;margin:0 0 0.5rem 0;">LitServe Tutorial: Fast and Easy AI Model Serving Framework</h3> <p style="font-size:0.875rem;color:#9ca0ab;margin:0;">Tutorial LitServe: Framework Serving Model AI yang Cepat dan Mudah Pendahuluan LitServe adalah framework open-source dar...</p> </div> </a> </div> </section> </div>  <noscript id="social-links-noscript"> <div style="display:none"> <a rel="me" href="https://www.facebook.com/rubythalibai">Facebook</a> <a rel="me" href="https://www.youtube.com/@rubyabdullah">YouTube</a> <a rel="me" href="https://www.instagram.com/rubythalib.ai/">Instagram</a> <a rel="me" href="https://www.linkedin.com/company/rubythalibai/">LinkedIn</a> <a rel="me" href="https://www.tiktok.com/@rubythalib.ai">TikTok</a> </div> </noscript>  <script> // Track all async operations window.domStability = { react: false, fonts: false, analytics: true, complete: false }; // Signal when everything is ready function checkDOMStability() { if (window.domStability.react && window.domStability.fonts && window.domStability.analytics) { window.domStability.complete = true; document.documentElement.setAttribute('data-dom-ready', 'true'); document.documentElement.setAttribute('data-page-loaded', 'complete'); window.dispatchEvent(new Event('dom-stable')); console.log('DOM is now stable and measurable'); } } // Make function globally available window.checkDOMStability = checkDOMStability; // Set timeout fallback to ensure signal fires even if scripts fail setTimeout(function() { window.domStability.react = true; window.domStability.fonts = true; checkDOMStability(); }, 3000); // 3 second max wait </script>  <script> window.dataLayer = window.dataLayer || []; function gtag(){dataLayer.push(arguments);} window.gaReady = false; function loadGA() { if (window.gaLoaded) return; window.gaLoaded = true; var script = document.createElement('script'); script.async = true; script.src = 'https://www.googletagmanager.com/gtag/js?id=G-6MKSY0HY48'; document.head.appendChild(script); script.onload = function() { gtag('js', new Date()); gtag('config', 'G-6MKSY0HY48'); window.gaReady = true; // Signal DOM stability after GA loads document.documentElement.setAttribute('data-ga-loaded', 'true'); // Trigger GTM loading after GA is ready setTimeout(function() { if (window.loadGTM) window.loadGTM(); }, 150); }; } var events = ['mousedown', 'touchstart', 'keydown']; var gaTimer = setTimeout(loadGA, 5000); events.forEach(function(event) { window.addEventListener(event, function() { clearTimeout(gaTimer); loadGA(); }, { once: true }); }); </script> </body> </html>