qforge-dev/torque

Torque

Torque is a declarative, fully typesafe DSL for quickly building complex LLM synthetic datasets. Compose conversations like components, generate realistic variations with any model efficiently.

✨ Features

🎯 Declarative DSL - Compose conversations like components
🔒 Fully Typesafe - Zod schemas with complete type inference
🔌 Provider Agnostic - Generate with any AI SDK provider (OpenAI, Anthropic, DeepSeek, vLLM, LLaMA.cpp etc.)
🤖 AI-Powered Content - Generate realistic varied datasets automatically without complicated scripts
🎭 Faker Integration - Built-in Faker.js with automatic seed synchronization for reproducible fake data
💰 Cache Optimized - Reuses context across generations to reduce costs
📉 Prompt Optimized - Concise, optimized structures, prompts and generation workflow lets you use smaller, cheaper models
♻️ Reusable Patterns - Build libraries of conversation templates
⚡ Concurrent Generation - Beautiful async CLI with real-time progress tracking while generating concurrently

🚀 Quick Example

import * as T from "@qforge/torque";
import { openai } from "@ai-sdk/openai";

await T.generateDataset(
  () => [
    T.generatedUser({ prompt: "Friendly greeting or introduction" }), // AI generated
    T.oneOf([
      // pick one randomly (weights are optional)
      { value: T.assistant({ content: "Hello!" }), weight: 0.3 }, // static
      T.generatedAssistant({
        prompt: "Respond to greeting",
        reasoning: T.generatedReasoning({
          prompt: "Reason about the greeting",
        }),
        // or reasoning: reasoning({ content: "...." }),
      }), // AI generated, gets remaining weight
    ]),
    T.times(between(1, 3), [
      T.generatedUser({
        prompt: "Chat about weather. Optionally mentioning previous message",
      }),
      T.generatedAssistant({ prompt: "Respond to user. Short and concise." }),
    ]),
  ],
  {
    count: 2, // number of examples
    model: openai("gpt-5-mini"), // any ai-sdk model
    seed: 42, // replayable RNG
    metadata: { example: "quick-start" }, // optional per-row metadata
  }
);

Outputs:

{"messages":[{"role":"user","content":[{"type":"text","text":"Hi there! I'm new here and just wanted to say hello."}]},{"role":"assistant","content":[{"type":"text","text":"Hello!"}]},{"role":"user","content":[{"type":"text","text":"The sunshine today is perfect for a walk in the park."}]},{"role":"assistant","content":[{"type":"text","text":"Absolutely—warm and bright out there."}]},{"role":"user","content":[{"type":"text","text":"Do you think the clouds will roll in later this evening?"}]},{"role":"assistant","content":[{"type":"text","text":"Maybe briefly, but it should stay mostly clear."}]}]}

{"messages":[{"role":"user","content":[{"type":"text","text":"Hey! Hope you're having a great day."}]},{"role":"assistant","content":[{"type":"text","text":"Hi there! I'm doing great—what can I help you with?"}]},{"role":"user","content":[{"type":"text","text":"The weather keeps flipping between sun and drizzle lately."}]},{"role":"assistant","content":[{"type":"text","text":"Totally—it’s been bouncing around all week."}]},{"role":"user","content":[{"type":"text","text":"Should I expect rain again tonight?"}]},{"role":"assistant","content":[{"type":"text","text":"Pack an umbrella just in case; there’s a chance of showers."}]},{"role":"user","content":[{"type":"text","text":"Thanks! I’ll be prepared if it turns stormy."}]},{"role":"assistant","content":[{"type":"text","text":"Good call—better to stay dry than sorry."}]}]}

💡 See full example: examples/quick-start.ts | ▶️ Try in Browser

🤔 Why Torque?

Building synthetic datasets for LLMs is tedious:

Sometimes you don’t have enough real data
Manual conversation writing doesn’t scale as conversations get long
Maintaining quality and consistency across thousands of examples is extremely time consuming
Tool calling patterns require intricate message sequences and are error‑prone
Generating different conversation flows means rewriting everything or creating various hard to maintain scripts
Designing generators that are random yet reproducible is surprisingly complex
Getting AI to understand complex composition scenarios (nested variations, conditional flows) takes significant prompt engineering time

Torque solves this with a declarative approach. Just like React transformed UI development from imperative DOM manipulation to composable components, Torque transforms dataset generation from manual JSON editing or writing complicated scripts to declarative conversation schemas. Plus, its optimized structure means you can use smaller, cheaper models while benefiting from cache optimization for lower costs.

📦 Installation

npm install @qforge/torque
# or
bun add @qforge/torque

📚 Core Concepts

Message Schemas

Build conversations by composing message schemas, you can compose them together to build complex conversations from reusable parts:

// Reusable greeting pattern
const greeting = () => [
  system({ content: "You are a helpful assistant." }),
  user({ content: "Hello!" }),
  assistant({ content: "Hi! How can I help?" }),
];

// Compose it with additional conversation
const extendedSchema = () => [
  ...greeting(),
  user({ content: "What's the weather like?" }),
  assistant({ content: "I'd be happy to check that for you!" }),
];

// Or create variations
const formalGreeting = () => [
  system({ content: "You are a professional assistant." }),
  user({ content: "Good morning." }),
  assistant({ content: "Good morning. How may I assist you today?" }),
];

const schema = () => [
  // Weighted selection between schema branches
  oneOf([
    { value: greeting(), weight: 0.6 },
    formalGreeting(),
    extendedSchema(),
  ]),
  // Continue with shared conversation flow
  generatedUser({ prompt: "Ask a question" }),
  generatedAssistant({ prompt: "Provide helpful answer" }),
];

💡 See full example: examples/schema-composition.ts | ▶️ Try in Browser

Row Metadata

Use metadata({ ... }) inside your schema to hoist custom fields into the generated row. The helper runs during the check phase, so you can safely compute metadata once and reuse it in generation. Passing a function receives the current metadata object (merged with any top-level metadata you provided) and may mutate it or return a new object for advanced scenarios like simple counters.

const schema = () => [
  system({ content: "You are a helpful assistant." }),
  oneOf([
    () => [metadata({ variant: "static" }), assistant({ content: "Hello!" })],
    () => [
      metadata({ variant: "generated" }),
      generatedAssistant({ prompt: "Greet the user warmly" }),
    ],
  ]),
];

const withCounter = () => [
  metadata((meta) => {
    meta.count = meta.count ?? 0;
    meta.count += 1;
    count += 1;
  }),
  generatedUser({ prompt: "Ask a question" }),
];

When the dataset is saved, you can read these values under row.meta.metadata.

Automatic ID Generation

When generating datasets with a seed, each row automatically receives a unique, deterministic id in its metadata. This ID is generated based on the seed value, making it easy to identify and track specific rows across multiple runs.

await generateDataset(schema, {
  count: 3,
  seed: 100,
  model: openai("gpt-4o-mini"),
});

// Output in row.meta.metadata:
// Row 0: { id: "row_100_1l2dpno" }  // seed: 100
// Row 1: { id: "row_101_txnff9" }   // seed: 101
// Row 2: { id: "row_102_2sx56u" }   // seed: 102

The ID combines the seed value with a deterministic hash, ensuring:

Reproducibility: Same seed always generates the same ID
Uniqueness: Different seeds produce different IDs
Traceability: Easy to reference specific examples in logs or when combining datasets

If custom metadata is provided, the ID is automatically merged with it:

await generateDataset(schema, {
  count: 2,
  seed: 100,
  model: openai("gpt-4o-mini"),
  metadata: { projectName: "my-project" },
});

// Output: { id: "row_100_1l2dpno", projectName: "my-project" }

Note: IDs are only generated when a seed is provided. Without a seed, no ID is added.

Composition Utilities

Build dynamic, varied datasets with composition helpers:

import { oneOf, times, between, optional } from "@qforge/torque";

const schema = () => [
  // Choose randomly from options (weights optional)
  oneOf([
    user({ content: "Hello" }),
    { weight: 0.5, value: user({ content: "Hi there" }) },
    user({ content: "Hey" }),
  ]),

  // Repeat pattern 3 times
  times(3, [
    generatedUser({ prompt: "Ask a question" }),
    generatedAssistant({ prompt: "Answer the question" }),
  ]),

  // Repeat random number of times (1-5)
  times(between(1, 5), [generatedUser({ prompt: "Follow-up question" })]),

  // Optionally include (50% chance)
  optional(assistant({ content: "Anything else I can help with?" })),
];

oneOf accepts plain schema entries or { value, weight } objects. Provide any subset of weights (summing to ≤ 1) and the remaining probability is spread evenly across unweighted entries.

Unique draws across a dataset

Pass a uniqueBy configuration when you need each option to be used at most once across every row/schema during generation. When using uniqueBy, each option must be an object with id, value, and optionally weight:

const toolOptions = [
  { id: "weather", value: weatherTool.toolFunction() },
  { id: "calendar", value: calendarTool.toolFunction() },
  { id: "flight", value: flightTool.toolFunction() },
];

const schema = () => [
  oneOf(toolOptions, {
    uniqueBy: {
      collection: "tools",
    },
  }),
];

You can also combine uniqueBy with weighted options:

const toolOptions = [
  { id: "weather", value: weatherTool.toolFunction(), weight: 0.5 },
  { id: "calendar", value: calendarTool.toolFunction(), weight: 0.3 },
  { id: "flight", value: flightTool.toolFunction(), weight: 0.2 },
];

const schema = () => [
  oneOf(toolOptions, {
    uniqueBy: {
      collection: "tools",
    },
  }),
];

The collection name identifies the shared pool (so multiple oneOf calls can coordinate). The id property must be a string, number, or boolean and is used to track uniqueness. Torque throws if the pool is exhausted, making it easy to guarantee perfect round-robin coverage.

Using `uniqueOneOf` factory function

For a simpler API, use uniqueOneOf to automatically generate IDs and create a reusable function. This is especially useful when you want to create the unique selection function outside of your schema:

import { uniqueOneOf } from "@qforge/torque";

// Create the factory function outside generation
const tools = [weatherTool, calendarTool, flightTool];
const oneOfTools = uniqueOneOf(tools);

// Or with weighted options
const weightedTools = [
  { value: weatherTool, weight: 0.5 },
  { value: calendarTool, weight: 0.3 },
  flightTool, // unweighted, gets remaining weight
];
const oneOfWeightedTools = uniqueOneOf(weightedTools);

const schema = () => {
  const tool = oneOfTools(); // Returns a unique tool each time
  return [
    tool.toolFunction(),
    generatedUser({ prompt: "Ask question requiring this tool" }),
    generatedToolCall(tool, "t1"),
    generatedToolCallResult(tool, "t1"),
  ];
};

The uniqueOneOf factory automatically:

Generates unique IDs for each item
Creates a unique collection name
Returns a function that enforces uniqueness across calls

💡 See weighted example: examples/weighted-one-of.ts
💡 Full utilities demo: examples/composition-utilities.ts | ▶️ Try in Browser

Tool Definitions

Define tools with Zod schemas for complete type safety:

import {
  tool,
  generatedToolCall,
  generatedToolCallResult,
} from "@qforge/torque";
import { z } from "zod";

// use standard tool schema using zod ensuring complete type safety
const weatherTool = tool({
  name: "get_weather",
  description: "Get current weather for a location",
  parameters: z.object({
    location: z.string().describe("City name"),
    units: z.enum(["C", "F"]).optional(),
  }),
  output: z.object({
    temperature: z.number(),
    condition: z.string(),
  }),
});

const schema = () => [
  weatherTool.toolFunction(),
  generatedUser({ prompt: "Ask about weather in a city" }),
  generatedToolCall(weatherTool, "t1"), // type safe 100% correct generated tool calls
  generatedToolCallResult(weatherTool, "t1"), // similarly 100% correct generated tool results
  generatedAssistant({ prompt: "Interpret the weather data for the user" }),
];

💡 See full example: examples/tool-calling.ts | ▶️ Try in Browser

🔐 TypeScript Support

Torque is built with TypeScript and provides complete type safety. Both for user and AI generating the data. Ensure that the arguments and tool results are always matching schema.

// Full type inference for tool parameters
const weatherTool = tool({
  name: "get_weather",
  description: "Get current weather for a location",
  parameters: z.object({
    location: z.string().describe("City name"),
    units: z.enum(["C", "F"]).optional(),
  }),
  output: z.object({
    temperature: z.number(),
    condition: z.string(),
  }),
});

// TypeScript knows the shape of parameters and output
weatherTool.toolCall("t1", {
  location: "NYC",
  units: "C", // ✅ Type-safe
  // units: 'K' // ❌ TypeScript error
});

weatherTool.toolCallResult("t1", {
  temp: 72,
  condition: "Sunny", // ✅ Type-safe
  // humidity: 50 // ❌ TypeScript error
});

Two-Phase Execution

Torque executes in two phases:

Check Phase - Analyzes conversation structure, registers tools
Generate Phase - Creates actual content with AI generation

This enables:

AI awareness of what are the exact steps in the conversation before generating content - you can create schemas where LLM "fills the gaps"
Accurate progress tracking
Pre-validation of conversation flow

Reproducible Generation with Seeds

Control randomness for reproducible datasets:

await generateDataset(schema, {
  count: 50,
  model: openai("gpt-5-mini"),
  output: "data/dataset.jsonl",
  seed: 12345, // Same seed = same output
});

How seeds work:

The seed parameter ensures deterministic generation across runs
Same seed + same schema = identical dataset structure everytime
Useful for debugging, testing, and versioning datasets
If omitted, a random seed is generated and displayed in the CLI
Seeds control both torque random selections and AI model sampling (when supported by the provider)

Background Token Counting

Token counts for each row are computed off the main thread using a worker pool so dataset generation stays responsive. Configure the pool with tokenCounterWorkers (default: 3), or disable counting entirely by setting it to 0.

await generateDataset(schema, {
  count: 20,
  model: openai("gpt-5-mini"),
  tokenCounterWorkers: 5, // spawn 5 token-counting workers
});

Output Formats

Choose your preferred output file format and data structure:

// Export as JSONL with default ai-sdk structure (default)
await generateDataset(schema, {
  count: 100,
  model: openai("gpt-4o-mini"),
  format: "jsonl",
  output: "data/dataset.jsonl",
});

// Export in OpenAI Chat Completions format (tools + messages structure)
await generateDataset(schema, {
  count: 100,
  model: openai("gpt-4o-mini"),
  format: "jsonl",
  exportFormat: "chat_template",
  output: "data/finetune.jsonl",
});

Supported File Formats (format):

jsonl (default) - JSON Lines format, one row per line. Best for streaming and line-by-line processing.
parquet - Apache Parquet columnar format. More efficient for large datasets and analytics tools (e.g., Pandas, DuckDB, Apache Spark).

Supported Data Structures (exportFormat):

ai-sdk (default) - Internal Torque format, compatible with Vercel AI SDK. Includes schema metadata, tool definitions, and full message objects.
chat_template - OpenAI Chat Completions compatible format. Flattened message structure with tools and messages top-level keys. Ideal for fine-tuning or direct API usage.

Both formats write rows incrementally as they're generated, so large datasets won't consume excessive memory.

💡 When format is specified without output, the file extension is automatically set based on the format.

💡 See full example: examples/parquet-export.ts

🔧 Advanced Examples

Async Tool Pattern

Model conversations where tools take time to execute:

import {
  generateDataset,
  generatedUser,
  generatedAssistant,
  generatedToolCall,
  generatedToolCallResult,
  tool,
  times,
  between,
} from "@qforge/torque";
import { z } from "zod";

const searchTool = tool({
  name: "web_search",
  description: "Search the web",
  parameters: z.object({ query: z.string() }),
  output: z.object({ results: z.array(z.string()) }),
});

await generateDataset(
  () => [
    searchTool.toolFunction(),

    // Initial request
    generatedUser({ prompt: "Ask for information requiring web search" }),

    // Tool call generated based on the user request
    generatedToolCall(searchTool, "search-1"),

    // Immediate acknowledgment
    searchTool.toolCallResult("search-1", "<tool_ack />"),

    generatedAssistant({
      prompt: "Acknowledge search started, assure user it's in progress",
    }),

    // Filler conversation while waiting.
    // While generating AI is aware how many messages are left.
    times(between(1, 3), [
      generatedUser({ prompt: "Casual conversation, unrelated to search" }),
      generatedAssistant({ prompt: "Respond naturally to casual topic" }),
    ]),

    // Actual result arrives with reused arguments
    generatedToolCall(searchTool, "search-1-FINAL", {
      reuseArgsFrom: "search-1",
    }),
    // Generated actual result based on previously generated tool call
    generatedToolCallResult(searchTool, "search-1-FINAL"),
    generatedAssistant({ prompt: "Present search results to user" }),
  ],
  {
    count: 50,
    model: openai("gpt-5-mini"),
    output: "data/async-tools.jsonl",
  }
);

💡 See full example: examples/async-tools.ts | ▶️ Try in Browser

Custom Generation Context

Guide the AI's generation style globally:

await generateDataset(schema, {
  count: 100,
  model: openai("gpt-5-mini"),
  output: "data/dataset.jsonl",
  generationContext: {
    global: {
      messages: [
        {
          role: "system",
          content:
            'Keep messages concise and natural. Avoid starting with "Sure" or "Thanks".',
        },
      ],
    },
    user: {
      messages: [
        {
          role: "system",
          content:
            "Generate diverse user messages with varying levels of technical detail.",
        },
      ],
    },
    assistant: {
      messages: [
        {
          role: "system",
          content:
            "Assistant should be helpful but concise. Use 2-3 sentences max.",
        },
      ],
    },
  },
});

💡 See full example: examples/custom-generation-context.ts | ▶️ Try in Browser

Multiple Tool Variations

Generate datasets with different tools:

import { oneOf } from "@qforge/torque";

const tools = [weatherTool, calculatorTool, searchTool];

await generateDataset(
  () => {
    const tool = oneOf(tools);

    return [
      tool.toolFunction(),
      generatedUser({ prompt: "Ask question requiring this tool" }),
      generatedToolCall(tool, "t1"),
      generatedToolCallResult(tool, "t1"),
      generatedAssistant({ prompt: "Present the result" }),
    ];
  },
  {
    count: 300, // 100 examples per tool
    model: openai("gpt-5-mini"),
    output: "data/multi-tool.jsonl",
  }
);

💡 See full example: examples/multiple-tool-variations.ts | ▶️ Try in Browser

Realistic Fake Data with Faker

Torque includes built-in Faker.js integration that automatically respects the seed system for reproducible fake data generation:

import {
  generateDataset,
  generatedUser,
  generatedAssistant,
  faker,
} from "@qforge/torque";

await generateDataset(
  () => [
    generatedUser({
      prompt: `Introduce yourself as ${faker.person.fullName()} from ${faker.location.city()}`,
    }),
    generatedAssistant({
      prompt: "Greet the user warmly",
    }),
  ],
  {
    count: 100,
    model: openai("gpt-5-mini"),
    output: "data/personas.jsonl",
    seed: 42, // Same seed = same fake names and cities
  }
);

Faker automatically uses Torque's seed system, so:

Same seed = identical fake data across runs
No manual seed configuration needed
Perfect for creating realistic user personas, product data, addresses, emails, etc.

Common use cases:

User personas: faker.person.fullName(), faker.person.jobTitle()
Locations: faker.location.city(), faker.location.country()
E-commerce: faker.commerce.productName(), faker.commerce.price()
Contact info: faker.internet.email(), faker.phone.number()
Dates: faker.date.future(), faker.date.past()

💡 See full example: examples/faker-integration.ts | ▶️ Try in Browser

🎨 CLI Features

Torque includes a beautiful CLI interface with:

Real-time progress bar showing completed/in-progress generations
Per-generation step tracking (e.g., "user message", "tool-call (web_search)")
Token counting for messages and tools
Concurrent execution with configurable workers
Seed display for reproducible runs
Output file location clearly shown

╭────────────────────────────────────────────────────╮
│ Dataset Generation                                 │
├────────────────────────────────────────────────────┤
│ Total:       100                                   │
│ Completed:   45                                    │
│ In Progress: 5                                     │
│ Seed:        42                                    │
│ Output:      data/dataset_2025-10-30.jsonl         │
│ Workers:     5                                     │
├────────────────────────────────────────────────────┤
│ ████████████░░░░░░░░░░░░░ 45%                      │
├────────────────────────────────────────────────────┤
│ #0: [████████████████░░░░] 80% tool-result (search)│
│ #1: [██████░░░░░░░░░░░░░░] 30% user message        │
│ #2: [████████████████████] 100% Writing...         │
│ #3: [██░░░░░░░░░░░░░░░░░░] 10% assistant message   │
│ #4: [██████████░░░░░░░░░░] 50% tool-call (calc)    │
╰────────────────────────────────────────────────────╯

🤝 Contributing

Contributions are welcome! This is part of a larger project exploring async tool patterns in LLMs.

📄 License

MIT License - see LICENSE for details

🔗 Related

Built with:

Vercel AI SDK - Universal AI provider interface
Zod - TypeScript-first schema validation
Bun - Fast JavaScript runtime

Made with ❤️ for the AI tinkerers community