liveBook · Manning

Seed dataset

As explained in LLM Customization and Fine-Tuning

A small set of initial examples used to condition or guide later synthetic generation.

A seed dataset is a small set of real examples used to guide synthetic data generation. Organize seeds by category so each generation prompt can use matching examples as style anchors. In a financial-sentiment workflow, 50–60 Financial PhraseBank examples from its AllAgree split provide the initial pool; prompts use two or three examples from the target category. section 3.6.1 section 3.6.2

Matching seeds help generated examples resemble the target data in sentence length, vocabulary, formality, and how sentiment is expressed. If the target category has too few seeds, the prompt can fall back to the full pool. section 3.6.2

How LLM Customization and Fine-Tuning covers it

How deep it goes
Defines seed selection with a 50–60-example pool and two-to-three-example prompts, but does not detail later generation steps.
Examples
Uses Financial PhraseBank's AllAgree split and category-matched financial sentiment sentences as style references.
What you can do after
Build category-conditioned generation prompts using real examples as style references.
How it connects
Connects synthetic data generation with few-shot prompting, sentiment classification, and distribution alignment.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of LLM Customization and Fine-Tuning. The section links point to the parts of the book it draws on.