Split View: Scale AI와 데이터 라벨링의 세계: AI 훈련 데이터 산업의 모든 것과 커리어 가이드

Scale AI Business Structure
├── Data Engine (Core Business)
│   ├── Image/Video Labeling (Autonomous Driving, Robotics)
│   ├── Text Labeling (NLP, LLM)
│   ├── 3D Point Cloud (LiDAR)
│   └── RLHF Data (LLM Alignment)
├── Government
│   ├── US Department of Defense Contracts
│   ├── Satellite Image Analysis
│   └── Intelligence Analysis Support
├── Generative AI Platform
│   ├── LLM Evaluation (Model Evaluation)
│   ├── Fine-tuning Data
│   └── Safety Data (Toxicity Classification)
└── Enterprise Solutions
    ├── Custom Pipelines
    ├── Quality Management Tools
    └── Analytics Dashboards

{
  "label": "car",
  "bbox": {
    "x_min": 120,
    "y_min": 80,
    "x_max": 350,
    "y_max": 240
  },
  "confidence": 0.95
}

# Panoptic Segmentation label example
panoptic_label = {
    "segments": [
        {"id": 1, "category": "road", "is_thing": False},      # stuff (background)
        {"id": 2, "category": "sky", "is_thing": False},        # stuff
        {"id": 3, "category": "car", "is_thing": True, "instance_id": 1},  # thing
        {"id": 4, "category": "car", "is_thing": True, "instance_id": 2},  # thing
        {"id": 5, "category": "person", "is_thing": True, "instance_id": 1}
    ]
}

"[Apple:ORG] CEO [Tim Cook:PERSON] announced a new product
 in [Cupertino:LOC]."

# 3D Bounding Box label
lidar_annotation = {
    "label": "vehicle",
    "center": {"x": 15.2, "y": -3.4, "z": 0.8},
    "dimensions": {"length": 4.5, "width": 1.8, "height": 1.5},
    "rotation": {"yaw": 0.35, "pitch": 0.0, "roll": 0.0},
    "num_points": 342,
    "tracking_id": "veh_0042",
    "attributes": {
        "vehicle_type": "sedan",
        "occlusion": "partial",
        "truncation": 0.0
    }
}

Prompt: "Explain quantum mechanics to an elementary school student"

Response A: "Quantum mechanics is about the rules of a very tiny world..."
Response B: "Quantum mechanics deals with particles smaller than atoms..."

Evaluation: A > B (Reason: Uses simpler analogies, age-appropriate vocabulary)

RLHF Pipeline: 5 Steps
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Step 1: Prompt Collection
  └─ Ensure diversity (topic, difficulty, language, culture)
  └─ Include safety test prompts
  └─ Include red-teaming prompts

Step 2: AI Response Generation
  └─ Generate multiple responses per prompt (typically 2-4)
  └─ Use different temperature/sampling settings
  └─ Can use different model versions

Step 3: Human Evaluation
  └─ Comparison: Choose A vs B
  └─ Rating: Score helpfulness, accuracy, safety separately
  └─ Correction: Edit directly to create "gold responses"

Step 4: Reward Model Training
  └─ Train reward function on human preference data
  └─ Based on Bradley-Terry model

Step 5: Policy Optimization
  └─ PPO (Proximal Policy Optimization) or
  └─ DPO (Direct Preference Optimization)

Scale AI Differentiators
━━━━━━━━━━━━━━━━━━━━━━━━

Strengths:
  + Largest network of skilled labelers
  + Government/defense security certification (FedRAMP)
  + Industry-best 3D point cloud labeling
  + Proven RLHF data pipeline
  + Automated quality management system

Weaknesses:
  - High pricing (burdensome for small teams)
  - Minimum contract size requirements
  - Limited self-service options
  - Customization takes time

# Labelbox Python SDK example
import labelbox as lb

client = lb.Client(api_key="YOUR_API_KEY")
project = client.create_project(name="Object Detection v2")

# Connect dataset
dataset = client.create_dataset(name="street_images_2025")

# Define ontology (label schema)
ontology_builder = lb.OntologyBuilder(
    tools=[
        lb.Tool(tool=lb.Tool.Type.BBOX, name="Vehicle"),
        lb.Tool(tool=lb.Tool.Type.BBOX, name="Pedestrian"),
        lb.Tool(tool=lb.Tool.Type.POLYGON, name="Road"),
        lb.Tool(tool=lb.Tool.Type.SEGMENTATION, name="Sidewalk"),
    ],
    classifications=[
        lb.Classification(
            class_type=lb.Classification.Type.RADIO,
            name="Weather",
            options=[
                lb.Option(value="sunny"),
                lb.Option(value="rainy"),
                lb.Option(value="cloudy"),
            ]
        )
    ]
)

from snorkel.labeling import labeling_function, PandasLFApplier
from snorkel.labeling.model import LabelModel

# Define labeling functions
@labeling_function()
def lf_keyword_positive(record):
    """Return POSITIVE if positive keywords are found"""
    positive_words = ["great", "excellent", "amazing", "love"]
    if any(w in record.text.lower() for w in positive_words):
        return 1  # POSITIVE
    return -1  # ABSTAIN

@labeling_function()
def lf_keyword_negative(record):
    """Return NEGATIVE if negative keywords are found"""
    negative_words = ["terrible", "awful", "hate", "worst"]
    if any(w in record.text.lower() for w in negative_words):
        return 0  # NEGATIVE
    return -1  # ABSTAIN

@labeling_function()
def lf_short_review(record):
    """Short reviews tend to be negative"""
    if len(record.text.split()) < 5:
        return 0  # NEGATIVE
    return -1  # ABSTAIN

# Combine noisy labels with Label Model
applier = PandasLFApplier(lfs=[
    lf_keyword_positive,
    lf_keyword_negative,
    lf_short_review
])
L_train = applier.apply(df_train)

label_model = LabelModel(cardinality=2, verbose=True)
label_model.fit(L_train, n_epochs=500, seed=42)
predictions = label_model.predict(L_train)

# Installation and launch
pip install label-studio
label-studio start

# Run with Docker
docker run -it -p 8080:8080 \
  -v label-studio-data:/label-studio/data \
  heartexlabs/label-studio:latest

class QualityMonitor:
    """Labeling quality monitoring system"""

    def __init__(self, golden_set_ratio=0.05):
        self.golden_set_ratio = golden_set_ratio
        self.annotator_scores = {}

    def inject_golden_items(self, task_batch, golden_items):
        """Randomly insert golden items into task batch"""
        import random
        n_golden = max(1, int(len(task_batch) * self.golden_set_ratio))
        selected_golden = random.sample(golden_items, min(n_golden, len(golden_items)))

        mixed_batch = task_batch.copy()
        for item in selected_golden:
            pos = random.randint(0, len(mixed_batch))
            mixed_batch.insert(pos, {**item, "_is_golden": True})
        return mixed_batch

    def evaluate_annotator(self, annotator_id, submissions):
        """Evaluate labeler accuracy on golden items"""
        golden_results = [s for s in submissions if s.get("_is_golden")]
        if not golden_results:
            return None

        correct = sum(
            1 for s in golden_results
            if s["submitted_label"] == s["golden_label"]
        )
        accuracy = correct / len(golden_results)
        self.annotator_scores[annotator_id] = accuracy

        if accuracy < 0.80:
            self._flag_for_retraining(annotator_id)
        return accuracy

from sklearn.metrics import cohen_kappa_score
import numpy as np

def compute_cohens_kappa(annotator1_labels, annotator2_labels):
    """Compute Cohen's Kappa between two annotators"""
    kappa = cohen_kappa_score(annotator1_labels, annotator2_labels)
    # Interpretation:
    # < 0.20: Poor/Slight agreement
    # 0.21-0.40: Fair agreement
    # 0.41-0.60: Moderate agreement
    # 0.61-0.80: Substantial agreement
    # 0.81-1.00: Almost Perfect agreement
    return kappa

def compute_fleiss_kappa(rating_matrix):
    """Compute Fleiss' Kappa for 3+ annotators"""
    n_items, n_categories = rating_matrix.shape
    n_raters = rating_matrix.sum(axis=1)[0]

    # Agreement per item
    p_i = (np.sum(rating_matrix ** 2, axis=1) - n_raters) / (n_raters * (n_raters - 1))
    p_bar = np.mean(p_i)

    # Chance agreement
    p_j = np.sum(rating_matrix, axis=0) / (n_items * n_raters)
    p_e = np.sum(p_j ** 2)

    # Fleiss' Kappa
    kappa = (p_bar - p_e) / (1 - p_e)
    return kappa

Quality Management Workflow
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

1. Assign same data to 3 labelers
2. All 3 agree -> Adopt that label
3. 2 agree, 1 disagrees -> Adopt majority + review dissenter
4. All 3 disagree -> Escalate to expert arbitrator
5. Repeated disagreement -> Signal to update guidelines

class AnomalyDetector:
    """System for detecting abnormal labeling patterns"""

    def detect_speed_anomaly(self, annotator_id, task_times):
        """Detect abnormally fast labeling (suspected random clicking)"""
        median_time = np.median(task_times)
        threshold = median_time * 0.3  # Less than 30% of median is suspicious

        suspicious_count = sum(1 for t in task_times if t < threshold)
        if suspicious_count / len(task_times) > 0.2:
            return {"status": "flagged", "reason": "speed_anomaly"}
        return {"status": "ok"}

    def detect_pattern_anomaly(self, annotator_id, labels):
        """Detect repeated same-label patterns"""
        from collections import Counter
        counter = Counter(labels)
        most_common_ratio = counter.most_common(1)[0][1] / len(labels)

        if most_common_ratio > 0.85:  # Over 85% same label
            return {"status": "flagged", "reason": "pattern_anomaly"}
        return {"status": "ok"}

class PreLabelingPipeline:
    """Pre-labeling pipeline"""

    def __init__(self, model, confidence_threshold=0.85):
        self.model = model
        self.confidence_threshold = confidence_threshold

    def pre_label(self, data_batch):
        """First-pass labeling with model, then route by confidence"""
        results = []
        for item in data_batch:
            prediction = self.model.predict(item)
            confidence = prediction["confidence"]

            if confidence >= self.confidence_threshold:
                # High confidence: auto-approve with sampling review
                results.append({
                    "item": item,
                    "label": prediction["label"],
                    "route": "auto_approve",
                    "confidence": confidence
                })
            elif confidence >= 0.5:
                # Medium confidence: human review (with pre-label reference)
                results.append({
                    "item": item,
                    "suggested_label": prediction["label"],
                    "route": "human_review",
                    "confidence": confidence
                })
            else:
                # Low confidence: human labels from scratch
                results.append({
                    "item": item,
                    "route": "human_label",
                    "confidence": confidence
                })
        return results

import numpy as np

class ActiveLearningSelector:
    """Active learning sample selector"""

    def uncertainty_sampling(self, model, unlabeled_pool, n_select=100):
        """Uncertainty-based sampling"""
        predictions = model.predict_proba(unlabeled_pool)
        # High entropy samples = most uncertain samples
        entropies = -np.sum(predictions * np.log(predictions + 1e-10), axis=1)
        top_indices = np.argsort(entropies)[-n_select:]
        return unlabeled_pool[top_indices]

    def diversity_sampling(self, embeddings, n_select=100):
        """Diversity-based sampling (samples far from cluster centers)"""
        from sklearn.cluster import KMeans
        kmeans = KMeans(n_clusters=n_select, random_state=42)
        kmeans.fit(embeddings)
        # Select the sample closest to each cluster center
        selected = []
        for i in range(n_select):
            cluster_mask = kmeans.labels_ == i
            cluster_points = embeddings[cluster_mask]
            distances = np.linalg.norm(
                cluster_points - kmeans.cluster_centers_[i], axis=1
            )
            selected.append(np.where(cluster_mask)[0][np.argmin(distances)])
        return selected

    def badge_sampling(self, model, unlabeled_pool, n_select=100):
        """BADGE: Combining uncertainty + diversity"""
        # Compute gradient embeddings, then select diverse samples via K-Means++
        gradients = self._compute_gradient_embeddings(model, unlabeled_pool)
        return self.diversity_sampling(gradients, n_select)

Data Flywheel Cycle
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

  More Data ──────→ Better Model
       ↑                    │
       │                    ↓
  More Users ←──── Better Service

Key: The more this cycle spins, the higher the barrier to entry (competitive moat)

# Example of an ML Data Engineer's daily work
# Automating labeling pipeline with Airflow DAG

from airflow import DAG
from airflow.operators.python import PythonOperator
from datetime import datetime, timedelta

default_args = {
    "owner": "ml-data-team",
    "retries": 2,
    "retry_delay": timedelta(minutes=5),
}

dag = DAG(
    "labeling_pipeline_v2",
    default_args=default_args,
    schedule_interval="@daily",
    start_date=datetime(2025, 1, 1),
    catchup=False,
)

def extract_raw_data(**kwargs):
    """Extract unlabeled data from S3"""
    pass

def run_pre_labeling(**kwargs):
    """Run pre-labeling with model"""
    pass

def distribute_to_annotators(**kwargs):
    """Distribute tasks to labeling platform"""
    pass

def quality_check(**kwargs):
    """Quality verification of completed labels"""
    pass

def export_training_data(**kwargs):
    """Export verified data as training dataset"""
    pass

extract = PythonOperator(task_id="extract", python_callable=extract_raw_data, dag=dag)
pre_label = PythonOperator(task_id="pre_label", python_callable=run_pre_labeling, dag=dag)
distribute = PythonOperator(task_id="distribute", python_callable=distribute_to_annotators, dag=dag)
qa = PythonOperator(task_id="quality_check", python_callable=quality_check, dag=dag)
export = PythonOperator(task_id="export", python_callable=export_training_data, dag=dag)

extract >> pre_label >> distribute >> qa >> export

Data Labeling Career Paths
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

[Entry Level]
  Annotation Specialist ($15-25/hr)
  │
  ├──→ Quality Reviewer ($25-40/hr)
  │     │
  │     └──→ Data Quality Manager ($70K-$120K)
  │           │
  │           └──→ Head of Data Operations ($150K+)
  │
  ├──→ RLHF Specialist ($80K-$150K)
  │     │
  │     └──→ AI Safety Researcher ($150K-$250K)
  │
  └──→ [Technical Transition]
        │
        ├──→ ML Data Engineer ($120K-$180K)
        │     │
        │     └──→ Senior ML Engineer ($180K-$250K)
        │
        └──→ Annotation Platform Engineer ($130K-$200K)
              │
              └──→ Engineering Manager ($200K+)

Determinants of AI Performance (Post-2025)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Algorithm differences:   ████░░░░░░ Diminishing
Computing power:         ██████░░░░ Can be solved with money
Data quantity:           ██████████ Nearly saturated (internet data limits)
Data quality:            ██████████ <- Ultimate bottleneck, requires human labor

Conclusion: Companies that can secure high-quality data will win the AI competition