2nd/10_Wiki/Topics/Coding/AI_Synthetic_Data.md

---
id: ai-synthetic-data
title: Synthetic Data — LLM 으로 train / test / fixture
category: Coding
status: draft
source_trust_level: B
verification_status: conceptual
created_at: 2026-05-09
updated_at: 2026-05-09
tags: [ai, synthetic-data, vibe-coding]
tech_stack: { language: "TS / Python", applicable_to: ["Backend"] }
applied_in: []
aliases: [synthetic data, LLM-generated data, test fixtures, data augmentation, anonymization]
---

# Synthetic Data

> LLM 가 fake data 생성. **Test fixture, ML training, 사용자 demo, anonymization**. Real data privacy / cost / scale 우회.

## 📖 핵심 개념
- Generation: LLM 가 schema 따라 data 생성.
- Augmentation: 기존 data 의 변형.
- Anonymization: PII 제거 + realistic 유지.
- Distillation: 큰 model → 작은 model 의 training.

## 💻 코드 패턴

### LLM 으로 fixture 생성
```ts
import { z } from 'zod';
import OpenAI from 'openai';
import { zodResponseFormat } from 'openai/helpers/zod';

const User = z.object({
  email: z.string().email(),
  name: z.string(),
  bio: z.string().max(200),
  interests: z.array(z.string()).max(5),
  age: z.number().int().min(18).max(80),
});

async function generateUsers(count: number): Promise<z.infer<typeof User>[]> {
  const r = await openai.beta.chat.completions.parse({
    model: 'gpt-4o-mini',
    messages: [
      { role: 'system', content: 'Generate diverse, realistic test user profiles. Vary demographics, names, bios.' },
      { role: 'user', content: `Generate ${count} users.` },
    ],
    response_format: zodResponseFormat(z.object({ users: z.array(User) }), 'users'),
  });
  return r.choices[0].message.parsed!.users;
}

const users = await generateUsers(50);
```

→ Faker.js 보다 realistic.

### Diverse generation
```ts
// 단순 — 비슷한 데이터 자주
// Better — diversity prompt

const prompts = [
  'Generate users from different countries',
  'Generate users with different age groups',
  'Generate users with different income levels',
];

const all: User[] = [];
for (const prompt of prompts) {
  const batch = await generateWithPrompt(prompt, 20);
  all.push(...batch);
}
```

### Schema-driven (any)
```ts
const Order = z.object({
  id: z.string().uuid(),
  userId: z.string().uuid(),
  items: z.array(z.object({
    productId: z.string().uuid(),
    quantity: z.number().int().positive(),
    price: z.number().positive(),
  })).min(1).max(10),
  status: z.enum(['pending', 'paid', 'shipped', 'delivered', 'cancelled']),
  createdAt: z.string().datetime(),
});

const orders = await generateFromSchema(Order, 100);
```

### Faker.js (deterministic, fast)
```ts
import { faker } from '@faker-js/faker';

faker.seed(42);  // deterministic

const user = {
  id: faker.string.uuid(),
  name: faker.person.fullName(),
  email: faker.internet.email(),
  address: {
    street: faker.location.streetAddress(),
    city: faker.location.city(),
    zip: faker.location.zipCode(),
  },
};
```

→ 빠름, 일관, but 패턴 명확 (LLM 보다 less realistic).

### Hybrid (Faker + LLM)
```ts
// Faker = structure (id, email, address)
// LLM = creative (bio, review text)

const user = {
  id: faker.string.uuid(),
  email: faker.internet.email(),
  bio: await llm.generate('Write a 100-character bio for a freelance designer'),
  reviews: await llm.generate('Write 3 realistic product reviews'),
};
```

### Test database seed
```ts
async function seed() {
  await db.user.deleteMany();
  await db.order.deleteMany();
  
  const users = await generateUsers(100);
  await db.user.createMany({ data: users });
  
  const orders = await generateOrders(500, users.map(u => u.id));
  await db.order.createMany({ data: orders });
  
  console.log(`Seeded ${users.length} users, ${orders.length} orders`);
}
```

```bash
yarn seed
```

→ Test environment 가 production-like.

### Anonymization (real → synthetic)
```ts
// Real user data → similar but anonymized
async function anonymize(user: User): Promise<User> {
  const r = await llm.complete({
    system: 'Generate a realistic user profile similar to this one but with all PII changed.',
    user: `Original: ${JSON.stringify(user)}`,
    response_format: { type: 'json_object' },
  });
  return JSON.parse(r);
}

// Or simpler — Faker
function anonymize(user: User): User {
  return {
    ...user,
    name: faker.person.fullName(),
    email: faker.internet.email(),
    phone: faker.phone.number(),
    // 비-PII keep (purchase history, preferences)
  };
}
```

→ Test on prod-like data without exposure.

### ML training data augmentation
```ts
// Few-shot examples → 더 많은 generation
async function augmentDataset(examples: Example[], targetSize: number) {
  const augmented: Example[] = [...examples];
  
  while (augmented.length < targetSize) {
    const batch = await llm.generate({
      system: 'Generate similar examples to these, with variations.',
      user: examples.slice(0, 5).map(e => JSON.stringify(e)).join('\n'),
      response_format: { type: 'json_object' },
    });
    augmented.push(...JSON.parse(batch).examples);
  }
  
  return augmented.slice(0, targetSize);
}
```

→ 100 examples → 1000.

### Distillation (big → small model)
```ts
// 1. Big model (GPT-4o) 가 답 생성
// 2. (input, output) 쌍 = training data
// 3. Small model (Llama 8B) fine-tune

async function generateTrainingData(inputs: string[]) {
  const data = [];
  for (const input of inputs) {
    const output = await openai.chat.completions.create({
      model: 'gpt-4o',
      messages: [{ role: 'user', content: input }],
    });
    data.push({ input, output: output.choices[0].message.content });
  }
  return data;
}

// 그 후 fine-tune small model.
```

→ Cost ↓ runtime, 비슷 quality.

### Edge case generation
```ts
async function generateEdgeCases(schema: any, count: number) {
  return await llm.generate({
    system: `Generate edge case test inputs based on this schema.
Include: empty, very long, special chars, boundary values, unicode, malformed.`,
    user: JSON.stringify(schema),
    response_format: { type: 'json_object' },
  });
}
```

### Adversarial (security test)
```ts
async function generateAdversarial(target: string, count: number) {
  return await llm.generate({
    system: `Generate adversarial inputs for security testing.
Include: SQL injection attempts, XSS, command injection, long strings, unicode tricks.`,
    user: `Target: ${target}`,
  });
}
```

→ Pen testing.

### Validation (synthetic 가 real 같은가?)
```ts
// Statistical check
const realStats = computeStats(realData);
const synthStats = computeStats(syntheticData);

// Distribution similarity (KS test, etc)
expect(ksDistance(realStats, synthStats)).toBeLessThan(0.1);
```

### Privacy guarantee
```
GDPR / HIPAA:
- Synthetic data 가 individual 추적 불가
- Differential privacy 가 강한 보장

Tools:
- gretel.ai
- Mostly AI
- YData
```

### Use cases
```
✅ Test fixtures (unit / integration / e2e)
✅ Demo / sandbox
✅ Load test data
✅ ML training augmentation
✅ Privacy-preserving sharing
✅ Edge case generation
✅ Adversarial testing

❌ Production data 대체 (real distribution 다름)
❌ Statistical analysis (bias)
```

### LLM-as-judge (synthetic 검증)
```ts
async function evaluateSynthetic(real: any[], synthetic: any[]) {
  return await llm.complete({
    user: `Compare these two datasets:
Real: ${JSON.stringify(real.slice(0, 10))}
Synthetic: ${JSON.stringify(synthetic.slice(0, 10))}

Are they similar in style, distribution, realism? Score 1-10. Output JSON.`,
    response_format: { type: 'json_object' },
  });
}
```

### Cost
```
1000 records × 100 tokens × $5/1M = $0.50

→ Cheap.

ML training data:
10K records × 500 tokens × $5/1M = $25

→ Still cheap vs human labeling.
```

### Reproducibility
```ts
// Seed
const seed = 42;
faker.seed(seed);

// LLM = non-deterministic. Use temperature 0 + cache.
const r = await openai.chat.completions.create({
  model: 'gpt-4o-mini',
  temperature: 0,
  seed: 42,  // 일부 model
  messages: [...],
});
```

### Volume
```ts
// 10K records — batch
const BATCH = 50;
const total = 10000;

const all: any[] = [];
for (let i = 0; i < total; i += BATCH) {
  const batch = await generate(BATCH);
  all.push(...batch);
  console.log(`${all.length}/${total}`);
}
```

→ Rate limit / cost 주의.

### Streaming (large dataset)
```ts
async function* generateStream(count: number) {
  for (let i = 0; i < count; i += 50) {
    const batch = await generate(Math.min(50, count - i));
    for (const item of batch) yield item;
  }
}

for await (const item of generateStream(10000)) {
  await db.insert(item);
}
```

### Tools
```
- Mockaroo (web): schema → CSV/JSON
- Faker.js / Faker (Python)
- gretel.ai: privacy-preserving synthetic
- SDV (Synthetic Data Vault): tabular ML
- LLM (GPT-4o, Claude, local)
```

### Best practices
```
1. Schema first (Zod / Pydantic)
2. Diverse prompts (variation)
3. Validation 가 real distribution 비슷
4. Privacy 검증 (no PII leak)
5. Versioning (synthetic dataset 도)
6. Cost monitoring
```

## 🤔 의사결정 기준
| 사용 | 추천 |
|---|---|
| Unit test | Faker (deterministic) |
| E2E test | Faker + LLM 조합 |
| Demo / sandbox | LLM (realistic) |
| ML training | LLM + augmentation |
| Privacy 보존 | gretel / Mostly AI |
| 큰 volume | Faker (cost) |

## ❌ 안티패턴
- **Real PII 변형 X — synthetic 가정**: privacy violation.
- **모든 거 LLM (큰 cost)**: Faker 가 OK 자주.
- **Distribution 가 real 같은 가정**: validate.
- **Reproducibility 없음**: test flake.
- **Seed 없음 (random)**: 다른 결과.
- **Edge case 없음**: 일반 case 만 generate.
- **Synthetic만 deploy production**: real 가 아님.

## 🤖 LLM 활용 힌트
- Schema-driven (Zod) + LLM = realistic.
- Faker (cheap) + LLM (creative) hybrid.
- Diverse prompt (multiple variation).
- Privacy-aware (no PII generation).

## 🔗 관련 문서
- [[Testing_Faker_and_Builders]]
- [[AI_Fine_Tuning_vs_Prompting]]
- [[AI_LLM_Eval_Patterns]]