Simula: A Framework for Designing Synthetic Datasets Based on Mechanisms and Reasoning
Google/DeepMind
Google Research introduces Simula, a framework that reframes synthetic data generation as dataset-level mechanism design. It uses reasoning-first methodology to construct entire datasets from first principles, enabling fine-grained control over coverage, complexity, and quality. Simula has been a key enabler for the Gemma ecosystem and safety classifiers at Google.
Google Research has introduced Simula, a framework for synthetic data generation that treats dataset creation as a mechanism design problem. Unlike existing methods that rely on manual prompts or seed data, Simula employs a reasoning-first approach, constructing datasets from first principles without seeds. It decomposes generation into four axes: deep taxonomies for coverage, local diversification for variety, complexification for difficulty adjustment, and quality checks via a dual-critic loop. Simula uses reasoning-based metrics like Taxonomic Coverage and Elo ratings for evaluation. It has been used with Gemini 2.5 Flash as a teacher and Gemma-3 4B as a student across domains like cybersecurity, legal reasoning, math (GSM8k), and multilingual knowledge (Global MMLU). Within Google, Simula powers the Gemma ecosystem (ShieldGemma, FunctionGemma, MedGemma), safety classifiers for Gemini, and features like scam detection for Android calls and spam filtering in Google Messages. The framework aims to make data generation a controllable science for specialized AI applications.
- Сокращения
- SQL = Structured Query Language
- LLM = Large Language Model
- TPM = Technical Program Manager
Source: Google Research —
original