Show HN: Torque – React for Datasets (usetorque.dev)

Torque is a new TypeScript toolkit that treats dataset generation like React: a declarative, typesafe DSL for composing conversation “components” and programmatically producing thousands of AI training examples. Users write reusable blueprints (e.g., generatedUser, generatedAssistant, oneOf) and call generateDataset with parameters like count, model, seed and metadata. Torque integrates Zod-powered typing and end-to-end type inference, streams progress in a CLI with concurrent workers, deterministic seeds, caching and replayable RNG, and supports any model reachable via common SDKs (OpenAI, Anthropic, DeepSeek, vLLM, LLaMA.cpp, etc.). It also adds observability—audit trails, token spend, policy checks, CI hooks, built-in evaluations and targeted retries—so runs are reproducible and debuggable. For the AI/ML community this addresses common scaling pains: brittle prompt scripts, QA burden, inconsistent examples, high token costs and complex conditional flows. By blending handcrafted prompts with AI-generated content, reusing context to lower token usage, and allowing smaller models + cache optimization, Torque aims to cut cost while keeping quality and repeatability. The key implications are faster iteration on training data, safer pipelines (typed schemas and policy gates), deterministic dataset regeneration for regression testing, and easier composition of complex multi-message tool-calling patterns—turning dataset engineering into maintainable, versioned code rather than ad-hoc spreadsheets.