skills.rest
Mercity-AI
Mercity-AIOFFICIAL
Explore all skills by this author →
GitHub

generate-dataset

Generates taxonomy-guided synthetic datasets as validated JSONL via OpenAI-compatible models.

4|Updated Jun 15, 2026
Search Tags:#llm#fine-tuning#data generation#synthetic data#json schema#taxonomy#dataset generation
One-click install
npx skills add https://github.com/Mercity-AI/Simula --skill generate-dataset-mercity-ai
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: generate-dataset
Source: https://github.com/Mercity-AI/Simula
Command: npx skills add https://github.com/Mercity-AI/Simula --skill generate-dataset-mercity-ai

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Creating high-quality synthetic training data requires more than prompting a model repeatedly — without a coverage plan, outputs cluster around the same patterns, and without validation, malformed or low-quality rows slip in. This Skill walks a user through the simula CLI pipeline (taxonomy → strategies → generation → critique → evaluation) while making the judgment calls about schema shape, variation axes, and representation policy so the user only answers two questions: what dataset and how much. ## Core Features & Use Cases - Guided end-to-end workflow: Classifies the task archetype (extraction, NL-query-to-structured, classification, QA, free-text), drafts the YAML config from the template, smoke-tests offline with fake models, then runs a paid pilot and scales up only after explicit user approval. - Distribution review before spending: Treats taxonomy.json and strategies.json as the dataset's distribution and mandates reviewing node weights, strategy bundles, and never_combine rules before any paid generation, preventing plausible-looking but skewed corpora. - Representation decision enforcement: Separates use-dependent choices (absent-field policy, label granularity, negative cases — asked with a recommendation) from universal quality rules (atomic fields, consistent formats, faithful content — enforced via schema, prompt modules, and the critic). - Use Case: A user asks for 5,000 job-posting extraction records. The Skill writes the config and a prompt module enforcing atomic fields and omit-when-absent policy, validates and smoke-tests offline, reviews the taxonomy and strategies, runs a 30-row pilot on real models, then scales to full size and reports accept rate and cost. ## Quick Start Ask the assistant to generate a synthetic dataset by describing what each row should contain, what it will be used for, and roughly how many rows you need.

Frequently Asked Questions about generate-dataset

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I generate a synthetic dataset with simula?▼

Describe the dataset and target row count, and the Skill writes a YAML config from the template, validates it, runs a free offline smoke test with fake models, then runs a small real pilot before scaling up. You only confirm before any paid model calls.

What is a taxonomy in synthetic data generation?▼

A taxonomy is a tree of variation factors (like topic, length, or tone) that conditions every generated record, forcing coverage across the space instead of near-duplicate outputs. Each row records its taxonomy_mix lineage so coverage can be audited.

Can I run simula without an API key?▼

Yes, setting every model role to "fake" runs the full pipeline offline and deterministically with no key. Real runs require the key named by provider.api_key_env in a gitignored .env file at the project root.

How do I enforce a fixed output schema for generated records?▼

Define a JSON Schema subset (objects, strings, numbers, arrays, enums, required fields) in the config; every record is validated against it and repaired or rejected on failure. For per-row varying schemas, use an open envelope plus a Python prompt module.

Why did my generated dataset come out skewed toward rare cases?▼

Skew usually comes from taxonomy node weights and strategy bundles: finely subdivided branches and strategies grouping unusual forms concentrate probability on them. Fix weights in taxonomy.json, rebalance strategies.json, and rerun — both files are reused verbatim.

What are the limitations of this dataset generation approach?▼

Generation is asyncio-concurrent and pilot-scale, not distributed, and only supports OpenAI-compatible chat endpoints. Lineage-based coverage reports what was requested, not verified content; use coverage_mode reassign for model-judged coverage at extra cost.

Related Skills

Explore More Data & Analytics Skills→
nirajrathod19nirajrathod19

design-exploration

Analyze components and generate a structured design brief for variation exploration.

Community
Advanced
Marwan78888Marwan78888
16

fine-tuning-expert

Configure LoRA/QLoRA fine-tuning workflows with JSONL datasets and hyperparameters.

Community
Advanced
ethan-rioethan-rio

arize-annotation

Manage human annotation workflows for Arize LLM observability projects.

Community
Intermediate
vignesh2027vignesh2027
6

synthetic-data-generator

Generate synthetic tabular, time-series, and image-like datasets with privacy controls.

Community
Advanced