pjt222/label-training-data

Trainingsdaten labeln

See Extended Examples for complete configuration files and templates.

Systematically label data for supervised ML with quality controls and efficient workflows.

Wann verwenden

• Starting supervised ML project that requires labeled training data
• Current model performance limited by insufficient labeled examples
• Need to label text, images, audio, or video data
• Want to measure and improve annotation quality
• Managing team of annotators with different expertise levels
• Implementing active learning to prioritize valuable examples
• Need to track labeling progress and costs
• Ensuring consistent labels across multiple annotators

Eingaben

• Erforderlich: Unlabeled dataset (images, text, audio, video)
• Erforderlich: Label schema (classes, attributes, or annotation types)
• Erforderlich: Labeling guidelines document
• Optional: Pre-existing labels (for quality comparison)
• Optional: Modellieren predictions for pre-annotation
• Optional: Budget and timeline constraints
• Optional: Domain expert availability for difficult examples

Vorgehensweise

Schritt 1: Installieren and Konfigurieren Label Studio

Einrichten Label Studio as the labeling platform.

# Install Label Studio
pip install label-studio

# Or use Docker for production
docker pull heartexlabs/label-studio:latest

# Create project directory
mkdir -p labeling-project/{data,exports,config}
cd labeling-project

# Initialize Label Studio
label-studio init my_project

# Start Label Studio server
label-studio start my_project --port 8080

Access at http://localhost:8080 (default Zugangsdaten: create on first visit).

For production deployment with Docker:

# docker-compose.yml
version: '3.8'

services:
  label-studio:
    image: heartexlabs/label-studio:latest
    ports:
      - "8080:8080"
# ... (see EXAMPLES.md for complete implementation)

docker-compose up -d

Erwartet: Label Studio running and accessible, PostgreSQL database initialized for production use.

Bei Fehler: If port 8080 already in use, change port in config, if Docker fails check Docker daemon is running, ensure sufficient disk space for data volumes, check firewall allows port 8080.

Schritt 2: Entwerfen Labeling Interface and Schema

Erstellen labeling configuration for your task type.

# labeling-project/config/labeling_config.py
"""
Label Studio configuration templates for common tasks.
"""

# Text Classification (single label)
TEXT_CLASSIFICATION = """
<View>
# ... (see EXAMPLES.md for complete implementation)

Erwartet: Labeling interface configured with appropriate controls for task type, data imported erfolgreich, interface accessible to annotators.

Bei Fehler: Validieren XML config with Label Studio's config validator, check data file format (JSON or CSV), ensure image/audio URLs are accessible if using external storage, verify API key has correct Berechtigungs.

Schritt 3: Vorbereiten Data and Implementieren Sampling Strategy

Format data for import and prioritize examples for labeling.

# labeling-project/prepare_data.py
import pandas as pd
import json
import random
from typing import List, Dict
from sklearn.cluster import KMeans
import numpy as np

# ... (see EXAMPLES.md for complete implementation)

Erwartet: Data formatted korrekt for Label Studio import, sampling strategy prioritizes informative examples, tasks include metadata for tracking.

Bei Fehler: Verifizieren JSON format with jq or Python json.load(), check that URLs are accessible if using remote images, ensure no special characters break JSON encoding, validate column names match config.

Schritt 4: Implementieren Quality Control and IAA Measurement

Einrichten processes to measure and improve annotation quality.

# labeling-project/quality_control.py
import pandas as pd
import numpy as np
from sklearn.metrics import cohen_kappa_score, confusion_matrix
from typing import Dict, List, Tuple
import logging

logging.basicConfig(level=logging.INFO)
# ... (see EXAMPLES.md for complete implementation)

Erwartet: Inter-annotator agreement measured (Cohen's Kappa > 0.6 is moderate, >0.8 is good), difficult tasks identified for review, annotator performance tracked.

Bei Fehler: If Kappa very low (<0.4), review labeling guidelines for clarity, retrain annotators, simplify label schema, check for ambiguous examples, consider using expert annotators for gold standard.

Schritt 5: Exportieren and Integrieren Labeled Data

Exportieren labels and prepare for ML training.

# labeling-project/export_labels.py
import requests
import pandas as pd
import json
from typing import List, Dict
import logging

logger = logging.getLogger(__name__)
# ... (see EXAMPLES.md for complete implementation)

Erwartet: Annotations exported in training-ready format, label distribution balanced or documented, data quality validated vor training.

Bei Fehler: Verifizieren API key Berechtigungs, check export format compatibility with your ML framework, handle missing annotations gracefully, validate JSON structure matches expected format.

Schritt 6: Set Up Continuous Labeling Pipeline

Automate labeling workflow with active learning integration.

# labeling-project/active_learning_pipeline.py
import schedule
import time
import logging
from datetime import datetime
from prepare_data import DataSampler, prepare_label_studio_format
from export_labels import LabelStudioExporter, convert_to_training_format
import pandas as pd
# ... (see EXAMPLES.md for complete implementation)

Erwartet: Active learning selects informative examples automatisch, labeling batches prepared weekly, model retrained when sufficient new labels available.

Bei Fehler: If uncertainty sampling doesn't improve model, try diversity sampling, if annotators can't keep up reduce batch size, monitor labeling queue length, implement backpressure if queue grows too large.

Validierung

• [ ] Label Studio accessible and responsive
• [ ] Labeling interface intuitive (test with sample annotator)
• [ ] Data import successful with correct format
• [ ] Inter-annotator agreement (Cohen's Kappa) > 0.6
• [ ] Quality control identifies problematic tasks
• [ ] Labels export in training-ready format
• [ ] Label distribution matches expected (or intentionally imbalanced)
• [ ] Active learning pipeline runs ohne manual intervention
• [ ] Annotation durchput meets project timeline

Haeufige Stolperfallen

• Unclear guidelines: Ambiguous instructions cause inconsistent labels; invest in detailed guidelines with examples
• Insufficient overlap: Can't measure IAA ohne multiple annotators per task; use 10-20% overlap
• Ignoring difficult cases: Edge cases often skipped but critical for model robustness; flag for expert review
• Batch effects: Annotator fatigue or learning causes temporal inconsistency; randomize task order
• No quality feedback: Annotators don't improve ohne feedback; provide regular accuracy reports
• Wrong sampling strategy: Random sampling wastes budget on easy examples; use uncertainty or diversity sampling
• Labeling in isolation: Domain experts needed for complex tasks; pair novices with experts initially
• Not tracking costs: Labeling expensive; monitor time per task and total budget consumption

Verwandte Skills

• version-ml-data - Version control for labeled datasets
• track-ml-experiments - Verfolgen model performance as labels added