2nd/10_Wiki/Topics/AI_and_ML/Synthetic-Data.md

---
id: wiki-2026-0508-synthetic-data
title: Synthetic Data
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Synthetic Data Generation, Synthetic Dataset, 합성 데이터, Artificial Data]
duplicate_of: none
source_trust_level: A
confidence_score: 0.95
verification_status: applied
tags: [synthetic-data, data-generation, llm, gan, diffusion, privacy]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: pytorch
---

# Synthetic Data

## 매 한 줄
> **"매 synthetic data 는 real data 의 statistical surrogate — privacy preserve + scale 의 unlock"**. 매 2026 LLM training 의 절반 이상 synthetic (Phi-4, Llama 4, Claude). 매 GAN→Diffusion→LLM-generated 의 evolution 의 끝. 매 validation gap 의 핵심 risk — 매 model collapse 의 prevent 의 첫 priority.

## 매 핵심

### 매 generation methods (2026)
- **LLM-augmented**: Self-Instruct, Evol-Instruct, magpie, persona-based generation. 매 dominant.
- **Diffusion (image/video)**: SDXL, FLUX, Sora-style. 매 image 의 standard.
- **GAN**: tabular (CTGAN), face (StyleGAN3) 의 niche only — 매 retire 진행.
- **Simulation**: Unreal/Unity, NVIDIA Omniverse — 매 robotics·AV 의 sim-to-real.
- **Rule/template**: Faker-style, structured format (JSON, SQL) — 매 reliable baseline.
- **Distillation**: teacher LLM → student dataset. 매 Phi-series approach.

### 매 use cases
- **LLM training**: instruction tuning, RLHF, code (Magicoder), math (MetaMathQA).
- **Privacy**: medical record (Synthea), financial (DPSDA differential privacy).
- **Robotics**: sim-to-real domain randomization, AV (Waymo Carcraft).
- **Edge cases**: rare disease, fraud — 매 real data 의 부족 area.
- **Augmentation**: minority class oversampling, MixUp.

### 매 validation (critical)
- **Fidelity**: marginal/joint distribution match (KS test, MMD, FID, KID).
- **Utility**: TSTR (Train Synthetic Test Real) — downstream metric.
- **Privacy**: membership inference, NN distance (DCR), k-anonymity check.
- **Diversity**: coverage, mode collapse detection.

### 매 model collapse
- **Definition**: synthetic-on-synthetic training 의 distribution narrow.
- **Mitigation**: real data anchor (Shumailov 2024 — 1% real / 99% synthetic 의 collapse 의 stop).
- **Provenance**: C2PA / watermark 의 future synthetic detection.

### 매 응용
1. **LLM instruction**: Self-Instruct + critic filter → 100k high-quality pairs.
2. **Tabular**: CTGAN / TVAE → DP-protected medical record.
3. **AV sim**: Carla / NVIDIA DRIVE Sim — millions of edge case km.
4. **Image augmentation**: SDXL controlnet → balanced classification dataset.

## 💻 패턴

### 1. LLM Self-Instruct (2026 magpie style)
```python
from anthropic import Anthropic
import json

client = Anthropic()

def magpie_generate(seed_topics, n_per_topic=20):
    """Magpie: prompt LLM with empty user → it generates instruction itself."""
    pairs = []
    for topic in seed_topics:
        for _ in range(n_per_topic):
            # First call: model invents user prompt
            user_msg = client.messages.create(
                model="claude-opus-4-7",
                max_tokens=200,
                messages=[{"role": "user", "content": f"Topic: {topic}\n\nGenerate one user question about this topic:"}],
            ).content[0].text

            # Second call: model answers it
            answer = client.messages.create(
                model="claude-opus-4-7",
                max_tokens=800,
                messages=[{"role": "user", "content": user_msg}],
            ).content[0].text

            pairs.append({"prompt": user_msg, "completion": answer})
    return pairs
```

### 2. Evol-Instruct (depth/breadth evolution)
```python
EVOLVE_PROMPT = """Rewrite the following instruction to make it more complex
(add constraints, deeper reasoning, edge cases). Output only the new instruction.

Original: {seed}
Evolved:"""

def evol(seed: str, rounds: int = 3):
    cur = seed
    for _ in range(rounds):
        cur = llm(EVOLVE_PROMPT.format(seed=cur))
    return cur
```

### 3. Critic filter (rejection sampling)
```python
JUDGE = """Rate this instruction-response pair 1-5 on:
- correctness, helpfulness, no hallucination.
Output JSON {"score": int, "reason": str}.

Q: {q}
A: {a}"""

def filter_pairs(pairs, threshold=4):
    keep = []
    for p in pairs:
        verdict = json.loads(llm(JUDGE.format(q=p["prompt"], a=p["completion"])))
        if verdict["score"] >= threshold:
            keep.append(p)
    return keep
```

### 4. CTGAN tabular synthesis
```python
from sdv.single_table import CTGANSynthesizer
from sdv.metadata import SingleTableMetadata
import pandas as pd

real = pd.read_csv("medical.csv")
meta = SingleTableMetadata()
meta.detect_from_dataframe(real)

syn = CTGANSynthesizer(meta, epochs=300, batch_size=500)
syn.fit(real)
fake = syn.sample(num_rows=10000)

# Quality check
from sdv.evaluation.single_table import evaluate_quality
report = evaluate_quality(real, fake, meta)
print(report.get_score())
```

### 5. Diffusion-based image synth (FLUX)
```python
import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16
).to("cuda")

prompts = [f"medical X-ray of {cond}, clear, anonymized" for cond in conditions]
images = pipe(prompts, num_inference_steps=20, guidance_scale=3.5).images
```

### 6. TSTR validation
```python
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.metrics import roc_auc_score

# Train on synthetic
clf = GradientBoostingClassifier()
clf.fit(syn_X, syn_y)

# Test on real held-out
auc = roc_auc_score(real_y_test, clf.predict_proba(real_X_test)[:, 1])
print(f"TSTR AUC: {auc:.3f}")  # close to TRTR baseline → high utility
```

### 7. Membership inference attack (privacy check)
```python
import numpy as np
from sklearn.neighbors import NearestNeighbors

def dcr_score(real, synthetic):
    """Distance to Closest Record — high = better privacy."""
    nn = NearestNeighbors(n_neighbors=1).fit(real)
    dists, _ = nn.kneighbors(synthetic)
    return np.mean(dists)
```

### 8. Real-data anchor (collapse prevention)
```python
def safe_mix(synthetic, real, real_ratio=0.1):
    """Shumailov 2024: small real anchor prevents collapse."""
    n_real = int(len(synthetic) * real_ratio / (1 - real_ratio))
    real_sample = real.sample(n=min(n_real, len(real)))
    return pd.concat([synthetic, real_sample]).sample(frac=1)
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| LLM instruction tuning | Magpie + Evol + critic filter |
| Tabular privacy | CTGAN + DP-SGD + DCR check |
| Image augment | FLUX/SDXL + controlnet |
| Robotics | Sim (Omniverse) + domain randomization |
| Fast structured | Faker / template |

**기본값**: LLM-generated + critic filter + real anchor (≥5%).

## 🔗 Graph
- 응용: [[Privacy-Preserving-ML]]
- Adjacent: [[Differential-Privacy]] · [[Model-Collapse]] · [[Data-Augmentation]]

## 🤖 LLM 활용
**언제**: instruction generation (Self-Instruct), critic judging, edge case ideation.
**언제 X**: privacy-sensitive numeric synth (LLM 의 number 의 hallucinate — CTGAN/DP method 사용).

## ❌ 안티패턴
- **Real data anchor 없 synthetic-only training**: 매 model collapse — distribution narrow.
- **Validation skip**: 매 unsafe deploy. TSTR / FID / DCR 의 minimum 3 metric.
- **Privacy claim without DP**: 매 pure synthetic ≠ private — membership inference 의 leak.
- **Single-method generation**: 매 mode-collapse risk. ensemble / diversity check.
- **Watermark / provenance 무시**: 매 future detection 의 impossible — C2PA 의 attach.

## 🧪 검증 / 중복
- Verified (Shumailov "AI models collapse" Nature 2024, Magpie paper 2024, Microsoft Phi-4 tech report 2025, NIST SP 800-188).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — synthetic data canonical (LLM-generated + GAN + diffusion + collapse) |