Core Concepts

The dataset generation system is built around four concepts: samples (the data unit), transforms (dataset generation pipeline) and datasets (collections of samples). Understanding these makes the rest of the docs easier to follow.

Sample

The fundamental unit of data. A sample contains:

Field	Description
seed	Raw starting data (news articles, documents, etc.)
question	Forecasting question generated from the seed
label	Ground truth answer with confidence score
prompt	Formatted prompt ready for model input
context	Additional context (news, RAG results)
meta	Custom metadata
is_valid	Whether the sample passed validation (default: True)

Samples flow through the pipeline: seeds produce questions, questions get labeled, and the result is a sample ready for training or analysis.

Transform

A dataset generation pipeline that processes data through multiple stages. Pipelines chain components:

• Seed generator — Produces raw data

• Question generator — Creates questions from seeds

• Labeler — Resolves questions with ground truth

• Context generators (optional) — Add relevant context

• Renderer (optional) — Format prompts

QuestionPipeline is the main orchestrator. You configure each stage and run it via lr.transforms.run(config).

When you run a pipeline, the SDK submits a job to the server. The job executes the pipeline stages and produces an output dataset.

Dataset

A collection of samples. Datasets serve as:

• Input — Feed custom samples into a pipeline via input_dataset

• Output — The result of lr.transforms.run() is a dataset

Datasets have an id and num_rows. Use dataset.download() or dataset.samples() to fetch samples, and dataset.flattened() for a flat list of dicts.