SYNTHETIC DATA
(insights from synthetic data primer)
Practitioners
Researchers
Big tech CEOS
Anshu Singh
Research Engineer
Data Practice, GovTech Singapore
5th Nov, 2025
Governments
About GovTechâs Data Practice
Areas spanning:
(The Launch) Upcoming Synthetic Data Primer ð
Bringing together recent advancement + practical perspectives
Strongly Positive Reviews
4
What you will learn today
5
What is
Synthetic Data?
Human Created Data = Real Data
Generated Data = âšSynthetic Data âš
Synthetic data is data that is not collected from real-world events
Train
Generate
Patterns, Correlations, Distributions and Relationships
Statistical and Generative AI models
(e.g., GANs, LLMs & Diffusion models)
Evaluate
Real Data
Synthetic Data
Focus on Broadly Accepted Definition Of Synthetic Data: Learning-Based Synthesis
statistical and generative synthesis methodologies that can learn and replicate the patterns and statistical properties of real data across text, images, tabular data, audio, and video.
Synthetic Data Generators - âSynthesizersâ
*Guidelines on selecting and benchmarking synthesizers are provided in the primer
Use Cases
+ Caveats
Synthetic Data is not Safe by Default
(1/4) Use Cases: Data Privacy
Real
Synthetic
Diminishing Returns
(2/4) Use Cases: Overcoming Data Scarcity And Accessibility
https://ifp.org/unlocking-a-million-times-more-data-for-ai/
by Andrew Trask (Google DeepMind)
Real
Synthetic
over a million times more data than what was used to train todayâs leading models.
Bias Overcorrection Risk
(3/4) Use Cases: Improving Data Representativeness
Real
Synthetic
Bias Overcorrection Risk
Avoid Synthetic Fantasies
(4/4) Use Cases: Simulating âWhat Ifâ Scenarios
time
time
Real
Synthetic
Avoid Synthetic Fantasies
âSelf-driving cars in Singapore now trained to detect cows sunbathing on car hoods blocking trafficâ
Quality And Privacy Evaluation
Quality Evaluation
Guidelines on selecting metrics for dimension evaluation:ￜ
flexibility
Fidelity
How similar?
Utility
How useful?
(proxy checks )
Privacy
How safe?
Generalization
Captures general patterns?
Fairness
How representative?
Integrity & Domain Coherence
Makes logical domain-specific sense?
Diversity & Coverage
Captures full variability?
conflicting
conflicting
fosters
fosters
fosters
conflicting
fosters
fosters
no universal definition of âgood qualityâ synthetic data
requires broader evaluation for practical, detailed and tailored assessment
common evaluation dimensions
Privacy Evaluation
Why Should You Care?
A synthesizer might learn more than just general patterns and end up leaking sensitive or identifying information through the synthetic data.
Context-driven privacy assessment
Attack-Based Evaluation: Adversarial Testing
Membership inference
Attribute inference
Record reconstruction
Determine whether a target individualâs record was used in training.
Infer a hidden/sensitive attribute about a target.
Recover an approximate training-set record (e.g., a near-duplicate).
increasing information gain for an attacker
Make realistic attacker assumption to avoid over-or underestimating risk
Assess privacy risks by testing how vulnerable or safe synthetic data is to privacy attacks.
Differential Private Synthetic Data Generation
Privacy guarantee: The synthetic data you see is almost as likely whether or not any single personâs row was in the training data.
(Google) Synthesizing and Adapting Error Correction Data for Mobile Large
Language Model Applications
Differentially Private Synthetic Data Deployments
From Governments and Enterprises
Researcher from Microsoft fine-tuned an LLM with DP on private data corpus. The model can be used to generate synthetic examples that resemble the private corpus
LLM-driven Synthetic Data Generation
Zero-Shot
One-Shot
Few-Shot
Chain-of-Thought
Multi-Step Decomposition
Multi-Agent
Knowledge Infusion
Self-Refinement
Controlled Prompting
PROMPTING
Full Fine-Tuning
Parameter-Efficient (LoRA, Adapters)
Instruction Tuning
Heuristic Checks
LLM-as-a-Judge
Re-weighting
Human Review
Auxiliary Models
Factualness
Relevance
Diversity
Privacy
Hallucination
Bias Amplification
Prompt Injection and Manipulation
Model Collapse
BASIC
ADVANCED
ENHANCEMENTS
FINE-TUNING
GENERATION
CURATION AND FILTERING
EVALUATION
RISK ASSESSMENT
Pipeline for optimal data generation
PROMPTING
FINE-TUNING
Approaches: Choice Depends on Use case & Resource Constraints
faster, cheaper and more flexible
Why use LLMs for Synthetic Data Generation?
Broad knowledge representation and pattern combination
In-context learning
Fine-grained controllability through prompting
(e.g., format specifications, constraints )
(thatâs a control panel :P)
Emergent reasoning
(e.g., multi-step problem solving, abstract reasoning )
Structured outputs
(e.g, JSON, table, code)
(But) Inherent bias and errors
Basic Prompting
Generate 10 synthetic citizen feedback entries for Singapore's
HealthHub digital services. Include variety in sentiment (positive,
neutral, negative) and feedback length. Make them realistic but
not based on real citizens.
Here's an example of a synthetic citizen service request:
{
"request_id": "MSF-2024-001",
"citizen_name": "Tan Wei Ming",
"age": 42,
"district": "Jurong West",
"service_type": "eldercare_assistance",
"request_details": "Seeking home care support for elderly parent",
"priority_level": "medium"
}
Generate 5 more synthetic citizen service requests following this exact format.
Here are examples of synthetic public officer feedback on digital transformation:
Example 1 (Positive, Tech Officer - GovTech):
"The new automated workflow system has reduced our processing
time by 40%. Excellent support from the technical team."
Example 2 (Constructive, Policy Officer - PMO):
"Digital tools are helpful, but we need better training for
senior officers. Perhaps structured upskilling programmes."
Example 3 (Mixed, Service Delivery - MSF):
"Citizens appreciate the online portal, but the backend integration
with legacy systems still causes delays in approvals."
Generate 8 more synthetic public officer feedback entries covering different agencies and transformation aspects.
Zero-shot Prompting
One-shot Prompting
Few-shot Prompting
Advanced Prompting (Requiring More Compute)
Generate synthetic policy recommendation documents for Singapore's Smart Nation initiative.
Think step by step:
1. First, identify a specific Smart Nation domain (transport, healthcare, urban planning, etc.)
2. Consider the current challenges in that domain
3. Propose evidence-based solutions
4. Structure as formal policy recommendation with rationale
Generate 3 policy recommendations following this reasoning process.
Create synthetic citizen journey data for Singapore's digital government services.
Step 1: Generate citizen profile (age, needs, digital literacy level)
Step 2: Based on profile, determine likely government services needed
Step 3: Map the digital touchpoints for each service
Step 4: Generate realistic interaction patterns and outcomes
Execute all steps for 5 different citizen personas.
Agent 1 (Generator): Create synthetic inter-ministry collaboration scenarios for whole-of-government initiatives.
Agent 2 (Validator): Review each scenario for:
- Realistic ministry roles and responsibilities
- Proper inter-agency coordination mechanisms
- Alignment with Singapore's governance structure
Agent 3 (Refiner): Enhance scenarios with specific details:
- Add relevant policy frameworks
- Include realistic timelines and resource allocation
- Ensure cultural and operational authenticity
Chain-of-Thought Prompting
Multi-Step Decomposition
Multi-Agent Prompting
Quality Enhancement
Task: Generate synthetic policy Q&A for housing policies
Step 1 - Retrieve: Get current HDB eligibility criteria from official sources
Step 2 - Context: Include retrieved facts in prompt
Step 3 - Generate: Create Q&A pairs based on verified information
"Question: What are the income requirements for first-time BTO applicants?
Answer: Based on current HDB guidelines, the monthly household income
ceiling is $14,000 for non-mature estates and $7,000 for mature estates..."
Generate diverse synthetic citizen complaints for Singapore government agencies.
Controlled Parameters:
- Agency: Rotate through [HDB, IRAS, LTA, NEA, PUB]
- Sentiment: [Frustrated, Constructive, Urgent, Appreciative]
- Complexity: [Simple query, Multi-step issue, Policy clarification]
- Channel: [Online form, Phone call, Walk-in, Email]
- Demographics: [Young professional, Family with children, Senior citizen, New resident]
Generate 20 complaints ensuring balanced distribution across all parameters.
Knowledge Infusion
(e.g., RAGs, knowledge graphs)
Controlled Prompting
(e.g., randomness, conditional scoping)
Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity. (2025)
Scaling Synthetic Data Creation with 1,000,000,000 Personas (2025)
BARE: Leveraging Base Language Models for Few-Shot Synthetic Data Generation (2025)
Evolving research into making prompting-based generation more diverse and higher quality
Quality Evaluation
(Focal Dimensions for LLM-based text generation)
Curation, Filtering & Evaluation
Risks
Model Collapse
Occurs when AI models are trained repeatedly on their own outputs, gradually losing fidelity and diversity
% of synthetic data in the training data
A Solution:
Diversify the data used by combining synthetic data with real data.
Misuse, Disinformation, and Misinformation Risks
Threatens both individual privacy and societal trust in information authenticity
TECHNOLOGY & OPERATIONS
HUMAN REVIEW
AI Governance Frameworks
Responsible And Ethical Use of Synthetic
Data ð±
ethical checks, standardised reporting etc.
watermarking, data provenance etc.
GOVERNANCE
Other Interesting Chapters
Outlook And Trends
âcan we generate better, more diverse (across modalities, domains, and scenarios), and safer data?â
generate cheaper, faster, and more realistic
combine or bridge modalities.
build trust and accountability
transparent, explainable, and collaboratve with human expertise
Practical Data Synthesis (Tabular Data)
Data Inception
Data Preparation
Data Pre-Processing
Synthesizer Selection & Training
Data Post-Processing and Curation
Data Evaluation
Deployment
Data Generation
âStakeholders reject utility metrics - privacy requirements stricter than expected
Poor quality distributions - need different synthesizer or encoding approach
Too many synthetic records got filtered; need to generate more synthetic data
https://forms.gle/mqrp8TRKidxcJxNN7
41
ORGANIZERS
https://responsible-synthetic-data.github.io/
https://responsible-synthetic-data.github.io/
INVITED SPEAKERS AND PANELISTS
(ACCEPTED PAPERS
TO BE RELEASED SOON)
Thank you.
Efficiency & Quality: Engineering Performance
Model progress
Cheaper LLM generation
Fine-tuning SLMs on the rise
Quality assurance
SDG cheaper, faster, and more realisticâwhile also pushing for stronger quality evaluation.