Overview

Training examples grounded in real outcomes

Dataset generation turns raw data into labeled forecasting samples through a configurable pipeline. You define the sources, question style, and labeling approach; the SDK handles the rest.

Pipeline Stages

A typical pipeline runs through these stages:

1. Seed generation — Fetch raw data (news articles, documents, etc.)

2. Question generation — Create forecasting questions from seeds using AI

3. Deduplication (optional) — Remove near-duplicate questions via exact or fuzzy matching

4. Context (optional) — Enrich samples with relevant news or RAG-retrieved documents

5. Labeling — Resolve questions with ground truth via web search

6. Rendering — Format questions into prompts for model input

7. Rollouts (optional) — Send prompts to one or more LLMs

8. Scoring (optional) — Score model outputs (from rollouts) against ground truth

QuestionPipeline

QuestionPipeline orchestrates these stages. You configure each stage and pass the config to lr.transforms.run():

import lightningrod as lr

client = lr.LightningRod(api_key="your-api-key")
binary_answer = lr.BinaryAnswerType()

# Get AI news to train a domain expert
seeds = lr.NewsSeedGenerator(
    start_date="2025-01-01",
    end_date="2025-04-01",
    search_query=["frontier AI model", "AI agents", "open source LLM", "AI research"],
)

# Define the scope and style of the questions
questioner = lr.ForwardLookingQuestionGenerator(
    instructions="Write forward-looking, self-contained questions with explicit dates/entities.",
    examples=["Will OpenAI publicly release GPT-5 by March 15, 2026?"],
    answer_type=binary_answer,
)

# Verify answers against live sources
labeler = lr.WebSearchLabeler(answer_type=binary_answer)

# Run pipeline
pipeline = lr.QuestionPipeline(seed_generator=seeds, question_generator=questioner, labeler=labeler)
dataset = client.transforms.run(pipeline, max_seeds=1000)

Running the Pipeline

• lr.transforms.run(config, input_dataset=None, max_seeds=None, max_cost_dollars=None, name=None, detach=False) — Submit and wait for completion. Returns a Dataset.

• lr.transforms.submit(config, input_dataset=None, max_seeds=None, max_cost_dollars=None, name=None) — Submit without waiting. Returns a TransformJob.

• lr.transforms.estimate_cost(config, max_seeds=None) — Estimate cost in dollars before running.

Use detach=True for long-running jobs so the job continues even if your local process exits.

Next Steps

• Core Concepts — Sample, Pipeline, Dataset, Transform Job

• Seed Generators — News, GDELT, FileSet, FileSetQuery

• Question Generators — Question, ForwardLooking, QuestionAndLabel, Template

• Deduplication — KeyDeduplication, fuzzy and exact field matching

• Labeling and Context — WebSearchLabeler, NewsContextGenerator

• Rollouts & Scoring — QuestionRenderer, RolloutGenerator, RolloutScorer, model consensus analysis

• Answer Types — Binary, Continuous, MultipleChoice, FreeResponse

• Datasets — Creating, fetching, and using datasets

• Examples — Notebooks and tutorials