Show HN: Torque – A declarative, typesafe DSL for LLM training datasets (MIT) (github.com)

Torque is a declarative, TypeScript-first DSL for building large, synthetic LLM datasets declaratively — think “React for conversation schemas.” It lets you compose reusable conversation components, inject AI-generated turns (generatedUser/generatedAssistant), and model tool usage with fully typesafe tool definitions via Zod. Torque is provider-agnostic (OpenAI, Anthropic, vLLM, LLaMA.cpp, etc.), includes Faker.js with seed-synced fake data, and exposes composition helpers (oneOf, times, between, optional) so complex, nested variations are simple to express and maintain. Technically, Torque runs a two-phase pipeline (Check Phase to validate structure and register tools; Generate Phase to produce content), supports deterministic seeds that control both schema randomness and model sampling, and offers cache optimization and context reuse to lower API costs and let you use smaller models. Tool calls, results and arguments are compile-time checked for shape, enabling safe generated tool interactions and reproducible async tool scenarios. The async CLI shows concurrent progress while generating, and generationContext lets you steer global/user/assistant style. For teams building scalable, reproducible synthetic datasets or complex tool-driven conversations, Torque reduces brittle scripting and prompt engineering while improving safety and maintainability through types and deterministic generation.